IL264507A

IL264507A - Systems and methods for transitioning patient care from signal-based monitoring to risk-based monitoring

Info

Publication number: IL264507A
Application number: IL264507A
Authority: IL
Original assignee: Etiometry Inc
Priority date: 2012-03-23
Filing date: 2019-01-28
Publication date: 2019-05-30
Also published as: JP2018077873A; IL277521A; IL234722B; JP6262200B2; EP2828780A1; EP2828780A4; IL264507B; IL277521B; DE202013012876U1; WO2013142268A1; JP6524196B2; JP2015512539A

Description

W0 2013/ 142268 PCT/US2013/031512 SYSTEMS AND METHODS FOR TRANSiTlONiNG PATlENT CARE FROM SiGNAL—BASED lViONlTORENG TO RISK-BASED ii/iONiTORlNG BACKGROUND The present disclosure relates to systems and methods for risk-—based patient monitoring. More particulariy, the present disciosure reiates to systems and methods for assessing the current and future risks of a patient by combining data of the patient from various different sources.

Practicing medicine is becoming increasingiy more compiicated due to the introduction of new sensors and treatments. As a resuit, clinicians are confronted with an avaianche of patient data, which needs to be evaluated and well understood in order to prescribe the optimai treatment from the muititucie of avaiiaoie options, while reducing patient risks. One environment where this avalanche of information has become increasingly problematic is the intensive Care Unit (lCU). There, the experience of the attending physician and the physician's ability to assimilate the available physioiogic information have a strong impact on the ciinical outcome. it has been determined that hospitais which do not maintain trained intensivists around the clock experience a 14.4% mortality rate as opposed to a 6.0% rate for fuiiy staffed centers. it is estimated that raising the ievei of care to that of average trained physicians across ali ICUS can save 160,000 lives and $4.38!‘: annually. As of 2012, there is a shortage of intensivists, and projections estimate the shortage wilt only worsen. reaching a level of 35% by 2020.

The value of experience in criticai care can be expieined by the fact that clinical data in the ICU is delivered at a rate far greater than even the most talented physician can absorb, and studies have shown that errors are six times more likely under conditions of information overioaci and eleven time more iikeiy with an acute time shortage. Moreover, treatment decisions in the iCU heavily rely on clinicai signs that are not directly measurabie, but are inferred from other physiologic information. Thus ciinician expertise and background play a more significant role in the minute to minute decision making process. Not surprisingly, this leads to a large variance in hidden parameter estimation. As an example, although numerous proxies for cardiac output are continuousiy monitored in critioai care, studies have demonstrated poor correlation between subjective assessment by clinicians, and objective measurement by therrnodilution. Experienced Eritensivists incorporate this inherent uncertainty in their decision process by effectiveiy conducting risk management, is. prescribing the treatment not only based on the most probabie patient state, but aiso weighing in the risks of the patient being in other more adverse states. From this perspective, experienced iritensivists confront the data U‘! WO 2013/142268 PCT/US2013/031512 overload in intensive care by converting the numerous heterogeneous signals from patient observations into a risk assessment.

Therefore, there is a clear need for a decision support system in the iCU that achieves a paradigm shift from signai—based patient monitoring to risk based patient monitoring, and consequentiy helps physicians overcome the barrage of data in the iC-U.

BREEF SUMMARY Disclosed herein is a risk-based patient monitoring system for criticei care patients that combines data from any of bedside monitors, eiectronic medicai records, and other patient specific information, to assess the current and the future risks to the patient. The system may be also embodied as a decision support system that prompts the user with specific actions according to a standardized medics! pian, when patient specific risks pass a predefined threshold. Yet another embodiment of the described technoiogies is an outpatient monitoring system which combines patient and famiiy evaluation, together with information about medication regiments and physician evaluations to produce a risk profile of the patient, continuously track its clinicai trajectory, and provide decision support to clinicians regarding when to schedule a visit or additionai tests.

According to one implementation, a risk based monitoring appiication executing on a system processor comprises a data reception module, a physiology observer rnoduie, a clinicai trajectory interpreter moduie, and a visuaiization and user interaction module. in an exempiary embodiment, the data reception module may be configured to receive data from bedside monitors, electronic medical records, treatment device, and any other information that may be deemed reievant to make informed assessment regarding the patients ciinical risks, and any combination thereof of the preceding elements.

The physioiogy observer module utiiizes muitiple measurements to estimate Probabiiity Density Functions (PDF) of Internet State Variables (iSVs) that describe the components of the physiology relevant to the patient treatment and condition. The ciinical traiectory interpreter rnoduie may be configured with multipie possible patient states, and determine which of those patient states are probable and with what probability, given the estimated probabiiity density functions of the internai state variabies. in various embodiments, the clinicai trajectory interpreter module determines the patient conditions under which a patient may be categorized and is capable of also determining the probable patient states under which the patient can be currentiy categorized, given the estimated probability density functions of the internai state variables. in this way, each of the (Ii.

W0 2013/142268 PCT/US2013/031512 possible patient states is assigned a probability value from 0 to ‘l. The combination of patient states and their probabilities is defined as the clinical risk to the patient.

The visualization and user interactions module takes i) time series of physiologic measurements acquired continuously or intermittently and patient specific identifiers such as condition, demographics, visual examinations from the data reception module; ii) time series of probability density functions of internal state variables estimated from the physiology observer module; and time series of the probabilities that the patient is at particular state and the hazard level of the respective risks from the clinical trajectory interpreter module. Then it visualizes this data on graphs which represent the dependence of the variables with time, by either directly plotting them on a screen, or in the case of ‘probability density functions plotting them by encoding the likelihood at particular point of time and at particular value with a color scheme.

The visualization and user interactions module may also visualize the current risks to the patient by representing them with boxes of different size and color, the size of the box corresponding to the probability of a patient state at particular point in time and the color of the box corresponding to its hazard level. Additionally, the visualization and user interactions module can allow the users to set alarms based on the patient state probabilities, share those alarms with other users, take notes related to the patient risks and share those notes with other users, and browse other elements of the patient medical history.

According to one aspect of the disclosure, as computer-implemented medium and method for risk based monitoring of patients comprises: A} acquiring, with a computer, data associated with a plurality of the internal state variables each describing a parameter physiologically relevant to one of a treatment and a condition of a patient; 8) storing, in a computer accessible memory, the acquired data associated with the plurality of the internal state variables; C) generating, with a computer, estimated probability density functions for the plurality of the internal state variables; and D) identifying, with a computer, from the generated probability density functions of the internal state variables, into which of a plurality of possible patient states the patient is currently categorizable and generating a probability value associated with each identified possible patient state. in one embodiment, the probability value associated with the identified possible patient states is between 0 and 1. In another embodiment, the method further comprises: E) presenting on a screen the probability values and their associated respective identified possible patient states, wherein the combination of identified possible patient states and their associated respective probability values is defined as the clinical risk to the patient.

WO 2013/142268 PCT/US2013/031512 According to another aspect of the disclosure, "a risk based monitoring system for monitoring patients comprises: a processor; a memory coupled to the processor; a data reception moduie, operably coupied to a plurality of sources of information relative to a patient, for acquiring data associated with a piuraiity of the internai state variables each describing a parameter physiologicaliy relevant to one of a treatment and a condition of a patient; a physiology observer module, in communication with the data reception module, and configured to generate probability density functions of the internal state variabies; a clinical trajectory interpreter module, in communication with the physiology observer module, and configured to identity into which of a piurality of possibie patient states the patient is currentiy categorizabie and to generate a probability value associated with each identified possible patient state. in one embodiment, the method further comprises: a user interaction moduie, in communication with the ciinicai trajectory interpreter and the data reception module and memory, for presenting the probabiiity values and their associated respective identified possible patient states, wherein the combination of identified possibie patient states and the associated respective probability values is defined as the clinical risk to the patient.

According to still other aspects of the disclosure, certain measurernents, such as Hemogiobin, are available to the system with an unknown amount of time latency, meaning the measurements are vaiid in the past relative to the current time and the time they arrive over the data communication links. The physioiogy observer module may handle such out of sequence measurements using back propagation, in which the current estimates of the lS\/s are projected back in time to the time of vaiidity of the measurements, so that the information from the latent measurement can be incorporated correctiy. Accordingiy, in accordance with another aspect of the disclosure, a computer-implemented method for risk based monitoring of patients, comprises: A) acquiring, with a computer, data associated with a piuraiity of the internal state variables each describing a parameter physioiogicaliy relevant to one of a treatment and a condition of a patient, not ail of the data associated with the piorality of the internal state variables with at the same periodicity; B) storing, in a computer accessibie memory, the acquired data associated with the plurality of the internal state variables; C) generating, with a computer, estimated probabiiity density functions for the plurality of the internal state variables; and D) identifying, with a computer, from the generated probability density functions of the Internet state variables, into which of a first piurality of possible patient states P(S1), P(S2), P(S3) P(S,,), the patient could has previoosiy been categcrizabie and generating a probability value associated with each identified possible prior patient state. in one embodiment, generating estimated probability density functions comprises: 01) generating estimated WO 2013/142268 PCT/US2013/031512 probability density functions for the first plurality of the internal state variables at a current time step ti; and C2) generating probability density functions for the plurality of the internal state variables at a another time step twp : where N is an integer value greater than 1, by evolving backwards from the probability estimates at time step tk to the time step tk_N using a defined transition probability kernel. l3RlEF DE.SCRiP"l'iON OF THE DRAWll\lGS it should be understood at the outset that although illustrative implementations of one or more embodiments of the present disclosure are provided below, the disclosed systems and/or methods may be implemented using any number of techniques, whether currently known or in existence. The disclosure should in no way be limited to the illustrative implementations, drawings, and techniques illustrated below, including the exemplary designs and implementations illustrated and described herein, but may be modified within the scope of the appended claims along with their full scope of equivalents. in the drawings: Figure 1 illustrates conceptually a medical care risk based monitoring environment in accordance with the disclosure; Figure 2A illustrates conceptually a basic schematic of the physiology observer module in accordance with the disclosure; Figures 28-D illustrate conceptually exemplary graphs of probability density functions for select ls‘./s as generated by the physiology observer module in accordance with the disclosure; Figure 3 illustrates conceptually a nomlimiting example of a physiology observer process in accordance with the disclosure; Figure 4 illustrates conceptually a non-limiting example of the physiology observer process in accordance with the disclosure; ' Figure 5 illustrates conceptually a time line, wherein back propagation is used to incorporate information in accordance with the disclosure; Figure 6 illustrates conceptually an example of a process involving mean arterial blood pressure (ABF’m) in accordance with the disclosure; Figure 7 illustrates conceptually an example of resampling in accordance with the disclosure; Figure 8 illustrates conceptually a clinical trajectory interpreter module using joined Probability Density Functions of iS\/s and performing state probability estimation to calculate the probabilities of different patient states in accordance with the disclosure; '19 W0 2013/ 142268 PCT/US2013/031512 Figure 9 illustrates conceptually a non~ilmiting example of a definition of a patient state employed by the clinical trajectory interpreter module in accordance with the disclosure; Figure 10 illustrates conceptually a non—limiting example of how a clinical trajectory interpreter module may employ the definition of patient states to assign probabilities that the patient may be classified under each of the four possible patient states at a particular point of time; Figure 11 illustrates conceptually an alternative approach of estimating the probabilities for different patient states in accordance with the disclosure; Figure 12 illustrates conceptually a non—limiting example of a definition of patient states assigned with hazard levels by the clinical trajectory interpreter module in accordance with the disclosure; ._ Figure 13 illustrates conceptually patient slates and their respective probabilities organized into tree graphs called etiologies in accordance with the disclosure; Figure 14 illustrates conceptually an exemplary etiology tree for a given set of patient states and physiologic variables in accordance with the disclosure; Figure ‘l5 illustrates conceptually a method for calculating the utility of different measurements in accordance with the disclosure; Figure 16 illustrates conceptually one possible realization of integration of external computation generated from third party algorithms in accordance with the disclosure; Figure 17 illustrates conceptually an example of integration instructions of an external computation in accordance with the disclosure; Figure ‘I8 illustrates conceptually an additional example of integration instructions of an external computation in accordance with the disclosure; Figure 19 illustrates conceptually example functionalities of the visualization and user interactions module in accordance with the disclosure; A Figure 20 illustrates conceptually an example of a summary View that may convey on. a single screen a risk profile for each patient in a particular hospital unit in accordance with the disclosure; Figure 21 illustrates conceptually one possible realization of a view describing the ongoing risks of the patient in accordance with the disclosure; Figure 22 illustrates conceptually how the slider on top of the patient View may be utilized in reviewing the history of the patient risks in accordance with the disclosure; W0 2013/142268 PCT/US2013/031512 Figure 23 illustrates how the user can navigate the etiology tree by clicking on the composite patient state and viewing the constituent patient states in accordance with the disciosure; I Figure 24 illustrates conceptually how in the same framework the user may view the predicted risks for the patient by siiding the slider ahead of current time in accordance with the ciisciosure; ' Figure 25 iiiustrates conceptually how the user may choose to set an alarm for a particular risk in accordance with the disciosure; Figure 26 illustrates conceptualiy yet another possible visualization of the patient risk trajectory, is, the evolution of the patient states’ probabiiities associated with particuiar risks in accordance with the disclosure; Figure 27 illustrates conceptually how the system may directiy visualize the probability density functions of various internai state variabies in accordance with the disclosure; Figure 28 iliustrates conceptually an example of a tagging feature that a user interface may impiement in accordance with the ciisciosure; Figure 29 illustrates conceptusliy a Newsfeed view of the user interface in accordance with the disclosure; Figure 30 iiiustrates conceptuaiiy a Condition Summary View of the user interface in accordance with the disclosure; Figure 31 illustrates conceptually the ability for the user interface to inciude and display reference materlai, which may be accessed through the internet, or stored within the system in accordance with the disclosure; Figure 32 illustrates conceptuaiiy a general Dynamic Bayesian Network (DEN) that may be employed to capture the physioiogy mode! of the HLHS stage t palliation patients in accordance with the disclosure; ’ V Figure 33 illustrates conceptually several equations that may be used to model the dynamics of the HLHS stage 1 physiology in accordance with the disclosure; Figure 34 iilustrates conceptually example equations that may be used to abstract the relationships between the dynamic variables in the model and the derived variables in accordance with the disciosure; Figure 35 illustrates conceptually a possibie observation model that may be used to relate the derived variabies with the available sensor data in accordance with the disclosure; W0 2013/142268 PCT/US2013/031512 Figure 36 illustrates conceptually possible attributes, patient states, and etiology tree that may be used by the clinical trajectory interpreter module in the case of the HLHS Stage 1 population in accordance with the disclosure; Figure 37 illustrates conceptually one possible environment in which the risk based monitoring system can be applied to assist clinicians in deciding whether to apply a particular treatment in accordance with the disclosure; Figure 38 illustrates conceptually a non-limiting example set of patient states relevant to blood transfusion that may be used to inform the blood transfusion decision in accordance with the disclosure; Figure 39 illustrates conceptually another application of the risk-based monitoring system, applying standardized medical plans in accordance with the disclosure; Figure 40 illustrates conceptually an example application of the risk based monitoring system combined with a specific type of standardized clinical plan in accordance with the disclosure; Figure 41 illustrates conceptually an example risk stratification that may be employed by the system in the context of Nitric Oxide treatment in accordance with the disclosure; Figure 42 illustrates conceptually possible patient states that may describe the clinical trajectory of an ADHD patient in accordance with the disclosure; Figure 43 lists the available patient evaluation modalities as M1, M2, and M3; Figure 44 illustrates conceptually a dynamic model of the patient evolution from state to state abstracted by a Dynamic Bayesian Network in accordance with the disclosure; Figure 45 illustrates conceptually an alternative embodiment for two predictions of how the patient state can transition in a single month given medication change or a dosage change in accordance with the disclosure; Figure 46 illustrates conceptually one possible embodiment and scenario of visualization displaying the patient clinical trajectory and risks in accordance with the disclosure; Fig. 47 illustrates conceptually an evaluation of the patient and the patient trajectory at week 9 at which point risk based patient monitoring system determines a probability distribution function for the state of the patient for each of the past nine weeks in accordance with the disclosure; Figure 48 shows a l’ollow—up evaluation based on teacher and parent Vanderbilt diagnosis in accordance with the disclosure; WO 2013/142268 PCT/US2013/031512 Figure 49 shows consequent evaluation based on all avaiiable measurements-office visit, parent and teacher evaluation, which establishes high probability for significant improvement in accordance with the disclosure; Figure 50 shows yet another follow-up at which point it is established that the patient is most probably stably improved, and has been stably improved between the two evaluations in accordance with the disclosure; Figure 51 shows a follow-up evaluation of the patient and the patient trajectory in the absence of measurements, wherein due to the lack of recent observation, the uncertainty is increasing in accordance with the disclosure; Figure 52 shows the state of this uncertainty given a full patient evaluation (all measurement modaiities) in accordance with the disclosure; and Figure 53 illustrates yet another possible visualization from the described system output. it shows possible patient state transitions under changes of treatment plan, e.g., change of medication in accordance with the disclosure.

DETAILED DESCRIPTION Technologies are provided herein for providing risl<—based patient monitoring of individual patients to clinical personnel. The technologies described herein can be embodied as a monitoring system for critical care, which combines data from various bedside monitors, electronic medical records, and other patient specific information to assess the current and the future risks to the patient. The technologies can be also embodied as a decision support system that prompts the user with specific actions according to a standardized medical plan, when patient specific risks pass a predefined threshoid. Yet another embodiment of the described technologies is an outpatient monitoring system which combines patient and family evaluation, together with information about medication regiments and physician evaluations to produce a risk profile of the patient, continuously track its clinical trajectory, and provide decision support to clinicians as regarding when to schedule a visit or additional tests.

System Modules And interaction Referring now to the Figures. Figure 1 illustrates a medical care risk based monitoring environment 1010 for providing health providers, such as physicians, nurses, or other medical care providers, risl<~based monitoring in accordance with various embodiments of the present disclosure. A patient 101 may be coupied to one or more physiological sensors or bedside monitors 102 that may monitor various physiological parameters of the patient. These physiological sensors may include but are not limited to, a blood oximeter, a blood pressure W0 2013/142268 PCT/US2013/031512 measurement device, a pulse measurement device, a giucose measuring device, one or more analyte measuring devices. an electrocardiogram recording device, amongst others. in addition, the patient may be administered routine exams and tests and the data stored in an electronic medical record (EMR) 103. The electronic medical record 103 may include but is not limited to stored information such as hemoglobin, arterial and venous oxygen content, lactic acid, weight, age, sex, ICD-9 code, capillary refill time, subjective ciinician observations, patient self-evaluations, prescribed medications, medications regiments, genetics, etc. in addition, the patient 101 may be coupled to one or more treatment devices 164 that are configured to administer treatments to the patient. in various embodiments, the treatments devices 104 may include extracorporeai membrane oxygenator, ventilator, medication infusion pumps, etc.

By way of the present disclosure, the patient 101 may be afforded improved risiotissed monitoring over existing methods. A patient specific risl<—based monitoring system, generaliy referred to herein as system 100, may be conﬁgured to receive patient related information, including real-time information from bed~side monitors 102, EMR patient information from electronic medical record 103, information from treatment devices 104, such as settings, infusion rates, types of medications, and other patient reiated information, which may include the patients medical history, previous treatment plans, results from previous and present lab work, ailergy information, predispositions to various conditions, and any other information that may be deemed relevant to make an informed assessment of the possible patient conditions and states, and their associated probabiiities. For the sake of simplicity, the various types of information listed above will generaliy be referred to hereinafter as “pstient«specific information". in addition, the system may be configured to utiiize the received information, determine the ciinical risks, which then can be presented to a medical care provider, including but not iimited to a physician, nurse, or other type of clinician.

The system, in various embodiments, includes one or more of the following: a processor 111, a memory 112 coupled to the process? 111, and a network interface 113 configured to enable the system to communicate with other devices over a network. in addition, the system may include a risk-based monitoring application 1020 that may include computer-executable instructions, which when executed by the processor 111, cause the system to be abie to afford risk based monitoring of the patients, such as the patient 101.

The risk based monitoring application 1020 inciudes, for example, a data reception module 121, a physiology observer module 122, a ciinicai trajectory interpreter module 123, and a visualization and user interaction module 124. in an exemplary embodiment, the data reception moduie 121 may be configured to receive data from bedside monitors 102, eiectronic WO 2013/142268 PCTfUS2013/031512 medical records 103, treatment devices 104, and any other information that may be deemed relevant to make an informed assessment regarding the patients clinical risks, and any combination thereof of the preceding elements.

The physiology observer module 122 utilizes multipie measurements to estimate probability density functions (PDF) of internal state variabies (ISV-5) that describe the components of the physiology relevant to the patient treatment and condition in accordance with a predefined physiology model. The i8Vs may be directly observabie with noise (as a non- limiting example, heart rate is a directiy observable iSV), hidden (as a non-limiting example, oxygen delivery (D02) defined as the flow of biood saturated oxygen through the aorta cannot be directiy measured and is thus hidden), or measured intermittently (as a non—lirniting exampie, hemoglobin concentration as measured from Cornpiete Blood Count tests is an intermittently observable lSV).

In one embodiment, instead of assuming that all variables can be estimated deterministicaiiy without error, the physioiogy observer module 122 of the present disclosure provides probability density functions as an output. Additional details related to the physioiogy observer module 122 are provided herein.

The ciinioal traiectory interpreter module 123 may be configured, for example, with multiple possible patient states, and may determine which of those patient states are probable and with what probabiiity, given the estimated probability density functions of the internal state variabies. A patient state is defined as a qualitative description of the physiology at a particular point of a clinical trajectory, which is recognizable by medicai practice, and may have implications to clinicai decisiommaking. Examples of particular patient states include, but are not limited to, hypotension with sinus tachycardia, hypoxia with myocardial depression, compensated circulatory shock, cardiac arrest, hemorrhage, amongst others. in addition, these patient states may be specific to a particular medical condition, andthe bounds of each of the patient states may be defined by threshoid values of various physioiogical variables and data. in various embodiments, the clinicai trajectory interpreter module 123 may determine the patient conditions under which a patient may be categorized using any of information gathered from reference materials, information provided by health care providers, other sources of information.

The reference materials may be stored in a database or other storage device 130 that is accessible to the risk based monitoring application 1020 via network interface 113, for example.

These reference materials may include material synthesized from reference books, medicai iiterature, surveys of experts, physician provided information, and any other material that may be used as a reference for providing medical care to patients. in some embodiments, the ‘ll 01 WO 2013/142268 PCTfUS2013/031512 ciinicai trajectory interpreter module 123 may first identify a patient population that is similar to the subject patient being monitored. By doing so, the clinical trajectory interpreter module 123 may be able to use relevant historicai data based on the identified patient population to help determine the possible patient states.

The clinical trajectory interpreter module 123 is capabie of also determining the probabie patient states under which the patient can be currentiy categorized, given the estimated probability density functions of the iniernai state variables, as provided by physioiogy observer module 122, In this way, each of the possibie patient states is assigned a probability value from O to 1. The combination of patient states and their prohabiiities is defined as the clinical risk to the patient. Additional detaiis related to the clinical traiectory interpreter moduie 123 are provided herein.

Visualization and user interactions module 124 may be equipped to take the outputs of the data reception module 121, the physiology observer moduie 122, and the clinical trajectory interpreter module 123 and present them to the ciinicei personnel. The visualization and user interactions module 124 may show the current patient risks, their evolution through time, the probability density functions of the internal state variabies as functions of time,‘ and other features that are caiculated by the two moduies 122 and 123 as by~products and are informative to medical practice. Actditioriaily, visuaiization and user interactions module 124 enables the users to set alarms based on the patient state probebiiities, share those alarms with other users, take notes related to the patient risks and share those notes with other users, and browse other elements of the patient medical history. Additionai details reiated to the visualization and user interactions moduie 124 are provided herein.

Physiology Observer Figure 2 illustrates a basic schematic of the physiology observer module 122, which utilizes two models of the patient physioiogy: a dynamic model 212 and an observation mociei 221. The dynamic model 212 captures the relationship arising between the internal state variables at some time tk and another ciose time tm, thereby enabiing modeling of the patient physioiogy as a system whose present state has information about the possible future evolutions of the system. Given the propensity of the patient physiology to remain at homeostasis through auto-reguiation, there is a ciear rationai of introducing such memory in internal state variables that are indicative of the homeostasis, e.g., oxygen deitvery and oxygen consumption.

The observation model 221 may capture the reistioriships between measured physiology variables and other internal state variaioies. Examples of such models iriciude: a) the 12 WO 2013/142268 PCT/US2013/031512 dependence of the difference between systolic and diastolic arterial blood pressures (also called pulse pressure) on the stroke volume; b) the relationship between heart rate, stroke volume, and cardiac output; :2) the relationship between hemoglobin concentration. cardiac output and oxygen delivery; d) the relationship between the Vanderbilt Assessment Scale and the clinical state of an attention deficit and hyperactivity disorder patient; and e) any other dependence between measurable and therefore observable parameters and internal state variables.

The physiology observer module 122 functions as a recursive filter by employing information from previous measurements to generate predictions of the internal state variables and the likelihood of probable future measurements and then comparing them with the most recently acquired measurements. Specifically the physiology observer module 122 utilizes the dynamic model 212 in the predict step or mode 210 and the observation model 221 in the update step or mode 220. During the prediction mode 210, the physiology observer module 122 takes the estimated PDFS of isvs 213 at a current time step tk and feeds them to the dynamic model 212, which produces predictions of the ISVS 211 for the next time step on. The is accomplished using the following equation: P(ISVS(t:c+-1)iM(tz.-)3 I P(15VS{€:m)ll’5VS(t;i))P(i5VS{€2c)lM(f;c} Dd-’5VS . ISVSEISV where ISVS(tr;)= {{SI/1(t;,),ISV2(tk),!$V3(tr),...I3V,.(tk) ] and M(tk) is the set of all measurements up to time tr. The probability P(ISVs(tk+1)iI5Vs(rk)) defines a transition probability kernel describing the dynamic model 212, which defines how the estimated PDFS evolve with time. The probabilities P(ISVs(tk)lM{rk)) are provided by the inference engine 222 and are the posterior probabilities of the ESVS given the measurements acquired at the previous time step. During the update mode 210 of the physiology observer module 122, the predicted lS\/s 211 are compared against the received measurements from data reception module 121 with the help of the observation model 221, and as a result the iSVs are updated to reflect the now available information. The inference engine 222 of module 122 achieves this update by using the predicted PDFs as a-priori probabilities, which are updated with the statistics of the measurements to achieve the posterior probabilities reflecting the current lSVs PDFS estimates 213. The inference engine 222 accomplished the update step 220 with the following equation which is Bayes’ Theorem, 13 WO 2013/142268 PCT/US2013/031512 {P .(m1‘{tk+1): ??_'l__2 (ta-.+1:’» mr>.(tk+_1)_li5{5k+__1))P (I5V${Fle+1).lM P(M(5k+1)) P(ISVs Where P(m1(t;¢+1),m2(tk+1),...m,,(tk+1)lISVs(tk+1)) is the conditional likelihood kernel provided by the observation model 221 that determines how likely the currently received measurements are given the currently predicted lSVs.

At the initialization time, e.g., i=0, when no current estimate of lSVs PDFs is available, the physiology observer module 122 may utilize initial estimates 250, which may be derived from an educated guess of possible values for the lS\/s or statistical analysis of previously collected patient data.

Figure 3 illustrates a non—limiting example of models that enable the physiology observer in accordance with the present disclosure. While not directly observable, the management of oxygen delivery, D02, is an important part of critical care. Therefore, precise estimation of D02 can inform improved clinical practice. in the illustrated example, this estimation is achieved through the measurements of hemoglobin concentration (l-lg), heart rate (HR), diastolic and systolic arterial blood pressures, and 8:30;‘. The dynamic model 212 assumes that oxygen delivery is driven by a feedback process which stabilizes it against stochastic disturbances.

Similarly, hemoglobin concentration is controlled around the norm value of 15 mg/dL.

The observation model 221 takes into account the relationship between arterial oxygen saturation SpO2, hemoglobin concentration and arterial oxygen content C302, the dependence of the difference between systolic, M3133, and diastolic, AEBPG, arterial blood pressures (also called pulse pressure) on the stroke volume, and the relationship between heart rate, HR, stroke volume, SV, and cardiac output. The two models are abstracted as a Dynamic Bayesian Network (DBN), and the physiology observer module 122 utilizes the DBN to continuously track the oxygen delivery. A Dynamic Bayesian Network is a systematic way to represent statistical dependencies in terms of a graph whose vertices signify variables (observable and unobservable), and whose edges show causal relationships. Further descriptions of an exemplary DBN for D02 estimation can be found in US. Provisional Application No. 6‘li699,492, filed on September 11, 2012, entitled SYSTEMS AND METHODS FOR EVAl_UATii\lG Ci_iNlCAi_ TRAJECTOREES AND TREATMENT STRATEGEES FOR OUTPATHENT CARE, Attorney Docket No. 44429—00‘iO7 PROV, and US. Provisional Application No. 6‘l,’684,24l, filed on August 1?, 2012, entitled SYSTEM AND METHODS FOR PROVlDlNG RlSl< ASSESSMENT lN ASSlS"i‘lNG CLlNlCiANS WITH EFFlClENT AND EFFECTNE BLOOD MANAGEMENT, Attorney Docket No. 44429-00108 PROV, to which 14 in WO 2013/142268 PCT/US2013/031512 priority is claimed, the disclosure of which is incorporated herein by reference.

Figure 4 depicts a non-limiting example of the physiology observer described above tracking 002, but over a ionger time intervai, i.e., 4 time steps. in the observer, the main hidden lSV is the oxygen delivery variable (D02). The two types of measurements, Hemoglobin (Hg) and oxirnetry {SpO2) are in dashed circles in Figure. 4. SpO2 is an example of the continuous or periodic measurements that the physioiogy observer module ‘£22 receives from sensors, such as bedside monitors 102 and treatment devices 104 connected to the patient 101 that continuously report information. Hemogiobin (Hg) is an exampie of an intermittent or aperiodic measurement extracted from patient lab work that is available to the observer on a sporadic and irregular basis, and iatent at times, relative to current system time. The physiology observer module 122 is capable of handling both types of measurements because, along with tracking the hidden iSVs, e.g. D02, module 122 also continuously maintains estimates of the observed vaiues for ail types of measurements, even when measurements are not present. Figure 4 depicts these estimates for the case of 85302 and Hg. As can be seen, the SpO2 measurements are available regularly at each time step, whereas Hg is only available at two of the time steps.

As mentioned above, certain measurements, such as Hemogiobin, are available to the system with an unknown amount of time latency. meaning the measurements are vaiid in the past relative to the current time and the time they arrive over the data communication ilnks. The physioiogy observer moduie 122 may handie such out of sequence measurements using back propagation, in which the current estimates of the iSVs are projected back in time to the time of validity of the measurements, so that the information from the latent measurement can be incorporated correctiy. Figure 5 depicts such time line. in Figure 5, hemoglobin arrives at the current system time, TR. but is vaiid and associated back to the ISV (002) at time 'i"k.9_. Back propagation is the method of updating the current lS‘\/s probabitity estimates P(ISVs(t;,)lM(tk)) with a measurement that is latent relative to the current time, rn(t.-,r¢_,,) Back propagation is accomplished in a similar manner to the prediction method described previously. There is a transition probability kernel, P(ISVs(t;,,_,,)l1SVs(tk)}, that defines how the current probabilities evolve backwards in time. This can then be used to compute probabilities of the iSVs at time t;(.,, given the current set of measurements which excludes the iatent measurement, as follows: P(ISVs(tk_,,)iM'(tk)) f;_,,,se,5,V i=(rsi'sn:,,_,,) §ISVs(tk))P(ISVs(tk) [M (ck) )d1SVs Once these probabilities are computed, the latent measurement information is incorporated using Bayes’ ruie in the standard update: ‘E5 WO 2013/142268 PCT/US2013/031512 P(m(tk-H)!ISV*§(.tk—7’1))P([SVS(tk*-H)iM(tk) ‘D(M'(tk)v7n(tk-71)) P U 51/'S(€k—n) i M (tic): m(hc—n)) ”~" The updated probabilities are then propagated back to the current time tk using the prediction step described earlier. Back propagation can be used to incorporate the information.

Another functionality of the physiology observer module 122 inciudes smoothing. The care provider using the system 100 may be interested in the patient state at some past time.

With smoothing, the physiology observer module 122 may provide a more accurate estimate of the patient lSVs at that time in the past by incorporating all of the new measurements that the system has received since that time, consequently providing a better estimate than the original filtered estimate of the overall patient state at that time to the user, computing P(I5'Vs(tk._.n)[M(tk)). This is accomplished using the first step of back propagation in which the probabiiity estimates at time ‘ck which incorporate all measurements up to that time are evolved backwards to the time of interest ta.“ using the defined transition probability kernel. This is also depicted in Figure 5, in which the user is interested in the patient state at 'Tk.,., and the estimates are smoothed back to that time.

Because physiology observer module 122 maintains estimates of each of the measurements available to the system 100 based on physiologic and statistical models, moduie 122 may filter artifacts of the measurements that are unrelated to the actual information contained in the measurements. This is performed by comparing the newly acquired measurements with the predicted likelihoods of probable measurements given the previous measurements. if the new measurements are considered highly unlikely by the model, they are not incorporated in the estimation. The process of comparing the measurements with their predicted likelihoods effectively filters artifacts and reduces noise. Figure 8 shows an example of such a process involving mean arterial blood pressure (Al3Pm). Because ABPm is collected using an intravenous catheter, the measured signals are often corrupted with artifacts that result in incorrect measurements when the catheter is used for medical procedures such as blood draws or line flushes. Figure 6 shows the raw ABPm measurements prior to being processed by the physiology observer with the measurement artifacts identiﬁed, as well as the filtered measurements after being processed by the physiology observer module 122. As can be seen, the measurement artifacts have been removed and the true signal is left. in various embodiments, physiology observer module 122 may utilize a number of algorithms for estimation or inference. Depending on the physiology model used, the physiology 16 WO 2013/142268 PCT/US2013/031512 observer module 122 may use exact inference schemes, such as the Junction Tree algorithm, or approximate inference schemes using Monte Carlo sampling such as a particle filter, or a Gaussian approximation algorithms such as a Kalman Filter or any of its variants.

As discussed, the physiology model used by physiology observer module 122 may be implemented using a probabilistic framework known as a Dynamic Bayesian Network, which graphically captures the causal and probabilistic relationship between the lSVs of the system, both at a single instance of time and over time. Because of the flexibility this type of model representation affords, the physiology observer module 122 may utilize a number of different inference algorithms. The choice of algorithm is dependent on the specifics of the physiology model used, the accuracy of the inference required by the application, and the computational resources available to the system. Used in this case, accuracy refers to whether or not an exact or approximate inference scheme is used. if the physiology observer model is of limited complexity, then an exact inference algorithm may be feasible to use. in other cases, for more complex physiology observer models, no closed form inference solution exists, or if one does exist, it is not computationally tractable given the available resources. In this case, an approximate inference scheme may be used.

The simplest case in which exact inference may be used, is when all of the isvs in the physiology model are continuous variables, and relationships between the lS\/s in the model are restricted to linear Gaussian relationships. In this case, a standard Kalman Filter algorithm can be used to perform the inference. With such algorithm, the probability density function over the lSVs is a multivariate Gaussian distribution and is represented with a mean and covariance matrix.

When all of the iSV‘s in the model are discrete variables, and the structure of the graph is restricted to a chain or tree, the physiology observer module 122 may use either a Forward- backward algorithm, or a Belief Propagation algorithm for inference, respectively. The Junction Tree algorithm is a generalization of these two algorithms that can be used regardless of the underlying graph structure, and so the physiology observer module 122 may also use this algorithm for inference. Junction Tree algorithm comes with additional computational costs that may not be acceptable for the application. in the case of discrete variables, the probability distribution functions can be represented in a tabular form. it should be noted that in the case where the model consists of only continuous variables with linear Gaussian relationships, these algorithms may also be used for inference, but since it can be shown that in this case these algorithms are equivalent to the Kaimen Filter, the lKairnan Filter is used -§{hc1]aS the example algorithm. 17 W0 2013/142268 PCT/US2013/031512 When the physiology model consists of both continuous and discrete lSVs with nonlinear relationships between the variables, no exact inference solution is possible. In this case, the physiology observer module 122 may use an approximate inference scheme that relies on sampling techniques. The simplest version of this type of algorithm is a Particle Filter algorithm, which uses Sequential importance Sampling. Markov Chain Monte Carlo (MCMC) Sampling methods may also be used for more efficient sampling. Given complex and non-linear physiologic relationships, this type of approximate inference scheme affords the most flexibility.

A person reasonably skilled in the relevant arts will recognize that the model and the inference schemes employed by the physiology observer module may be any combination of the above described or include other equivalent modeling and inference techniques.

When using particle filtering methods, a resamoling scheme is necessary to avoid particle degeneracy. The physiology observer may utilize an adaptive resampling scheme. As described in detail below, regions of the iSV state space may be associated with different patient states, and different levels of hazard to the patient, The higher the number, the more hazardous that particular condition is to the patients health. in order to ensure accurate estimation of the probability of a particular patient condition, it may be necessary to have sufficient number of sampled particles in the region. it may be most important to maintain accurate estimates of the probability of regions with high hazard level and so the adaptive resampling approach guarantees sufficient particles ‘will be sampled in high hazard regions of the state space. Figure 7 illustrates an example of this resampiing. State 1 and State 2 have the highest hazard level. The left plot depicts the samples generated from the standard resarnpiing.

Notice there are naturally more particles in state 1 and state 2 region because these states are most probable. The right plot shows the impact of the adaptive resampling. Notice how the number of samples in the areas of highest risk has increased significantly.

Clinical trajectory interpreter Referring now to Figure 8, the Clinical Trajectory interpreter 123 takes the joint Probability Density Functions of the lsvs from physiology observer module 122, and performs state probability estimation 801 to calculate the probabilities of different patient states. The Probability Density Functions of the i8\/'s may be defined in closed form, for example multidimensional Gaussians 260, or approximated by histogram 280 of particles 2?’O, as illustrated in Figures 2843. in both cases, the probability density functions of the lSVs can be referred to as: P(ISV1(t),ISV2(t),...,ISV-n(t)), where i is the time they refer to. Given the internal state variables the patient state may be defined by a conditional probability density function: 18 WO 2013/142268 PCT/US2013/031512 P(S| ISV1,5l»’2,...,ISVn.), where S ES1,$2,...,SN represents all possibie patient states 8; Then determining the probability of the patient being in a particular state 3, may be performed by the equation: P(s,(i)) = f_°°w... ff; Prsi ISV1, 15:72, , I31/,,) P(I$V1(t), [Si/2(6), , isi;,n)) dISV1 dist/;,_ in case that P(ISV1(t),ISV2(t),...,ISVn{t)) is defined by a closed form function such as multidimensional Gaussian 260, the integration may be performed directiy. in case that P(ISV1(t),ISV2(t),...,ISVn(t)) is approximated by a histogram 280 of particles 270 and P(S'l I5V.,,ISi/2, ISV,,) is defined by a partition of the space spanned by ISV1,.lSl/2, ...,ISVn into regions as shown in Figure 9, the probabiiity P(.'§,-(t)) may be calculated by calculating the fraction of particles 270 in each region.

Once patient state probabilities are estimated, the clinical trajectory interpreter module 123 may assign different hazard ievels 802 for each patient state or organize the states into different etiologies 803. The ciinicai trajectory interpreter module 123, in conjunction with the physiology observer module 122, may perform measurements utility determination 804 to determine the utility of different invasive measurements such as invasive blood pressures or invasive oxygen saturation monitoring. in one embodiment, the Clinical trajectory interpreter Module 123 determines the probebiiities that the patient is in a particular state, rather than the exaot state that the patient is in.

Figure 9 illustrates a non~limiting exsmpie of a definition of a patient state that may be employed by the clinical trajectory interpreter module ‘I23. Specifioaliy, it assumes that the function P(S[ [Si/1_,ISVz,...,ISi/,,) may be deﬁned by partitioning the domain spanned by the internal state variables1SV1,ISi/2,...,ISV;,. The particuiar example assumes that the patient physiology is described by two internal state varialaies: Pulmonary Vascular Resistance (PVR) and Cardiac Output (CO). The particular risks and respective etiologies that may be captured by these two lSVs emanate from the effects of increased pulmonary vascuiar resistance on the circulation. Speoiiicaliy, high PVR may cause right—heart failure and consequently reduced cardiac output. Therefore, P\/R can be used to define the attributes of Normal PVR and High PVR, and C0 to define the attributes of Normal CO and Low CO, by assigning threshoids with the two variables. By combining these attributed, four separate states can be defined: State 1: ‘i9 U‘: W0 2013/142268 PCT/US2013/031512 Low (30, Normal PVR; State 2: Low CO, High PVR; State 3: Normal CO, High PVR; State 4: Normal CO Normal PVR.

Figure 10 illustrates a non—limiting example of how the clinical trajectory interpreter module 123 may employ the definition of patient states to assign probabilities that the patient may be classified under each of the four possible patient states at a particular point of time. in the example, the clinical trajectory interpreter module 123 takes the joint probability density function of P(Cardiac Output (Tk), Pulmonary Vascular Resistance (Tk)) and integrates it over the regions corresponding to each particular state, which produces P(S‘l(Tl<)), P(S2(‘l”lt)), P(S3(Tk)), and P(S4(‘i'k)). in this way, the clinical trajectory interpreter module 123 assigns a probability that a particular patient state is ongoing, given the information provided by the physiology observer module 122. Note that if the output of the physiology observer module 122 is not a closed form function 260 but a histogram 280 of particles 270, the clinical interpreter will not perform integration but just calculate the relative fraction of particles 270 within each region.

Figure 11 illustrates an alternative approach of estimating the probabilities for different patient states. in this alternative approach, to calculate the probabilities F’(S1), P(S2), H83) and P(S4), the clinical trajectory interpreter module 123 employs the joint probability functions of the |SVs for two consecutive time windows TR and Tm to calculate a moving window average. Note in the example that the size of the window is doubled for two time instances, which indicates that the window may be of an arbitrary, suitable size. As a result of this moving window averaging, the clinical trajectory interpreter module 123 performs a dynamic analysis of the trajectory of the lS\/s. That is, it gives a metric of the probability that the physiology trajectory, as described by the lSVs, may be found in a particular region in a particular time frame. in other words, this probability calculation gives an estimate of the probability that a particular patient state may be ongoing in the chosen tlme~frame, as opposed to just at a chosen time instance.

Clinical trajectory interpreter module 123 may also assign hazard levels to each particular state. Figure 12 illustrates a non-limiting example of a definition of patient states assigned with hazard levels by the clinical trajectory interpreter module 123. The hazard levels may be informed from clinician surveys, reference literature or any other clinical sources. in the particular example, the clinical trajectory interpreter module 123 distinguishes between four different hazard levels: 1 —— Minimal risk, 2 — Mild risk, 3 — Medium risk, and 4 — Severe risk. The combination of the probability of a patient state and its hazard level will be referred from hereon as a “Patient risk." Figure 13 illustrates how the patient slates and their respective probabilities may be organized into tree graphs called etiologies. In particular, the attributes normal and low WO 2013/142268 PCT/US2013/031512 associated with the cardiac output lSV are the base nodes of the graph. Each of these vertexes has two children associated with the attributes of the pulmonary vascuiar resistance. This organization leads to each patient state being a leaf (end vertex) on the tree. This particular tree will be referred to as an etiology tree. The etiology tree may be further empioyecl by the visualization and user interaction module 124 to provide a layered view of the various patient risks as further described herein.

Figure 14 illustrates that the etiology tree may not be unique for a given set of patient states and physiologic variables. Speclﬁcaiiy, Figure 14 provides an alternative etiology tree for the example from Figure 13. The root of the alternative etiology tree starts from the attributes associated with the pulmonary vascular resistance, instead of the attributes associated with cardiac output. it can be appreciated that different ruies may be employed for generating the trees depending on various factors and the context of use. For example, one etiology tree may be preferred against another reeiization in different clinical situations or depending on the preference of the users. Moreover, the tree may ciynamicaliy change as the risks change and the clinical situation evolves.

Utility of Different Measurements During hospital care, there exist measurements that may harm the patient or slow down their recovery. Examples of such harmful measurements are ail measurements coming from catheters such as invasive blood pressures and blood oximetry, which have been shown to significantly increase the risk of infection. Therefore, it may be useful if, during the care process. the clinician is provided with an assessment of the utility of each of the potentially harmfui measurements. Figure 15 illustrates a method for calculating the utility of different measurements.

Referring to Figure 15, the risk-based system 100 and the clinical traiectory interpreter module 123 may calculate the utility of a particular measurement with the illustrated procedure.

Particularly. in step 9001, a measurement rn. may be selected. Given measurement mi and a current time (tcurrent)u in step 9002, the clinical trajectory interpreter module 123 may submit an instruction to the physiology observer module 122 to simulate the physiology observer module output (the probability density functions of the internal state variables) from a given arbitrary point back from the current time (tc.,,,em-T) to the current time imam with removal of measurement mi from the algorithm output. Then, in step 9003, the clinical trajectory interpreter module 123 may simulate the state probabilities estimation given the simulated output of the physiology observer and arrive with a set of patient state probabilities, i.e., Ps;m(S1(tcu,,em}), 21 Z0 Z5 WO 2013/142268 PCT/USZGI3/031512 Ps;,,,(S2(tc,ma,,r)),..., Ps;m(S,,(tCm,,t)). Then, in step 9004, using the state probabilities determined from all available measurements, i.e., P(S»,(tw,,..3m)), P(S2{t¢,,,em)),..., P(Sn(tc.,.,,.em)), the clinical trajectory interpreter module 123 may calculate the utility of the measurement m; using the formula: _. . 2 _. _ . . P(5‘(? 0) U(mi) —‘ D(Psim':P) — Z;n=1 P(Si(tCiH'l‘ent))}0g(Psm1ES1§:::::ent)))a which is also the Kuilback-Leibler divergence between the patient state distribution given all available measurements and the patient state distribution given the measurement mi has been removed for a time interval T.

Alternatively, in step 9005, the clinical trajectory interpreter module 123 may caiculate utility for m-, by employing the hazard levels. ri, assigned to each state Si by the formula: . _ ‘ _ __ __ _ . P(5l(t 'rre Q) U([nl:} — Dweighted(Ps1m|P) " 2&1 riP(S:(tcurrent))]08 P:h'n"‘(‘S“]‘E':‘;ur:e:‘“t)))- In a similar manner, the clinical trajectory interpreter module 123 can perform the utility calculation not only for a particular measurement, but also for any group of measurements. The utility calculation can also include a component that captures the potential harm associated with a particular measurement. For example, the invasive catheter measurement described above would have a large level of harm associated with it. in this way, the calculation trades the harm associated with the measurement against the vaiue of information it provides. An example of this modified utility calculation is given by the following formula: U(mi) "7 Dweightedwsiniw) “‘ H(mi)i where H(m.) defines a function that describes the harm of each available measure.

The risk-based monitoring system 100 can also integrate externai computation generated from third party algorithms implemented either on the same computation medium as the patlent~based monitoring system or as a part of an external device. Figure 16 illustrates one possible realization of integration of an external computation generated from third party algorithms. Particularly, the output from the external computation 9110 is provided to the clinical trajectory interpreter module 123 which impiernents integration instructions 9120. As a result the state probability estimation 801 produces new states P(newS1), P(newS2), P(newS3) 22 WO 2013/142268 PCTfUS2013/031512 P(newS,,+,,,), which may result in an increased number of states n+m from the original number of it states. Simiiariy, integration instructions 9120 may be provided to the hazard level assignment 802 and the etiology organization 803.

Figure 17' illustrates an example of integration instructions. in the example, it is assumed that the external computation, EC, provides information about particular binary attributes A-'=a1 or A=a;, and the specific of how the provided information is captured in the integration instructions by the conditional probability P(EC|A). Aiso, given four original states 81, S2. S3, and S4, the integration instruction may specify how the states 83 and 84 may be updated with two additional attributes A=a1 and A=a2, and turn into four new states nevi/S3, newS4 newss, and newS6. To perform this update, the integration instruction may also employ prior probabilities P(AlS5) and P(AiS4). These prior probabilities may be derived from retrospective studies by V analyzing what fractions of patients exhibiting state 83 or 84 have concomitantly exhibited A"-=a1 or A=a2.

Another way to derive the prior probabilities is by soiiciting the opinion of clinicians.

By utilizing the integration instructions, the state probabilities estimation 801 of the new states may then be derived from the formula: F’(A=aJ, S; E E0) = P(EC l A=aj) P(A=aJlS.} P(S,-) I P(EC), where i in {3,4} and j in {1,2}, and where P(S;) are the original patieni state probabilities derived from the output of the physiology observer module 122.

Figure 18 illustrates an additional example of integration instructions of an external computation. Again, the risk—based monitoring system 100 can perform the integration, as shown in Figure 18, both in the case that the external computation 9110 is generated on the some computational medium as the patiermbasecl monitoring system, or as a part of an external: device. in this case, it is assumed that the external computation 9110 provides direct information about a particular internal state variabie estimated by the physiology observer module 122 (or enhanced physiology observer module 9300). Therefore, to integrate the externai computation, the physiology observer module 122 can treat the external computation 9110 as an additional measurement and integrate it directly into the observation model 221. i;7asuéi_iz‘aii'sii anti33;Jser..;i.ni:éi§¢iio}-igiriiééj Figure 19 illustrates example functionaiiiies of the visualization and user interactions module 124. Specifioaliy, module 124 may receive all available patient information and data 23 1.5 WO 2013/142268 PCT/US2013/031512 including, the data from the data reception module 121, the joint probability density function produced by the physiology observer module 122, and the etiology tree, the risks, and the invasive measurements utilities estimated by the clinical trajectory interpreter module 123. By utiiizing this information, the visualization and user interactions module 124 may produce: 1) a unit view 1501 of patients describing their risks, diagnoses, etc; 2) a View 1502 of a patients electronic medical record inciuding laboratory results, prescribed medication, diagnoses, etc; 3) a view 1503 of a patients ongoing risks; 4) a view 1504 of a patients risk trajectory, i.e., how the probabilities for particular patient states have evolved in a particular time frame; 5) a View 1505 of a patients measurements utilities in the estimation of the particular patient risks; 6) a plot of a patient estimated lS\/s’ PDFs 1506 describing the time evolution of the lSVs’. PDFS; 7} a View 150? enabling to navigate through the etiology tree of the patient and thus visuaiize different levels of the tree; 8) a view 1508 showing a patients predicted risks; 0) a view 1509 enabling clinicians to view and set patient risk based alarms; 10) a view 1510 of a patients physiology monitoring data and its evolution against time; and 11) any combination of the above described. in addition, the visualization and user interactions moduie 124 may also produce a patient tags view 1511, a patient condition summary 1512, reference and training material 1513, and annotation and tag setting ‘E514.

Figure 20 illustrates an example of a summary view 2000 that may convey in a single screen a risk profile for each patient in a particular hospital unit. The risk profile represents what is the cumulative probability of the patient being in a particular hazard level. it is calouiatecl by summing the current probabilities of ail states at particular hazard level. in the exampie, the summed probabilities, hazard levels are represented by the height of four bars, each bar corresponding to a particular hazard level. in this specific example, these hazard levels may be Green (slanted hatching) — Minimal risk, Yeilow (vertical hatching) —- Mild risk, Orange (horizontal hatching) -» Medium risk, and Red (dotted hatching) — Severe risk.

Figure 21 illustrates one possible realization of a view 2100 describing the ongoing risks of the patient. Each round-cornered box corresponds to a particular risk: the coior corresponds to the hazard levei with Green (sianteci hatching) -— Minimal risk, Yellow (vertical hatching) —~ ll/iiicl risk, Orange (horizontal hatching) — Medium risk, and Red (dotted hatching) ~ Severe risk; the height of the box corresponds to the probability of the particular patient state. Risks are grouped in columns based on their hazard levels. The screen and the respective risks are updated in real~ttme as new data becomes available.

Still referring to Figure 21, in addition to the visualization of the ongoing patient risks, the system 100 may provide information about the utility of the various invasive measurements in 24 WO 2013/142268 PCT/US2013/031512 determining these risks. Specifically, the illustrated example gives the utilities of invasive arterial blood pressure (ABP) and invasive central venous pressure (CVP) measurements. The utility may be represented by filled bars 2110 and 2120, and the maximum utility may correspond to six filled bars. The six filled here may be displayed in color gradient from 1—dark green, 2~light green, 3—yellow, 4~red, 5-purple, to 6—white or empty. in this particular embodiment, the filled bars 2110 for ABP show all six colors, while two of the filled bars 2120 for CVP respectfully show 1—darl< green and 2~lig ht green and the remaining filled bars 2110 show 6-white or empty.

Figure 22 illustrates a View 2200 and how a slider 2210 or other graphic element on top of the patient view may be utilized in reviewing the history of the patient risks. Specifically, in the example, the slider 2210 is moved to show the patient risks at approximately four hours back from current time. This enables clinicians to review the continuous evolution of the patient risks and compare them with the applied treatment or any other external factors. In various embodiments, slider 2210 may be moved on the user interface with a pointing device, a command, or, if utilized in conjunction with touch sensitive displays, through touching and dragging the slider or other graphic element to designate the desired time period.

Referring now to both Figures 21 and 22, the etiology tree is used to combine the two states State A 2139: Hypoxia with low cardiac output and State B 2140: hypoxia with low Qp:Qs from Figure 21 to represent them by a single patient state (Hypoxia) 2220. In the particular example, this is used to fit the text into the smaller box of Figure 22 relative to Figure 21. The user can navigate the etiology tree in View 2300 by clicking on the composite patient state 2220 and viewing its constituent patient states 2130 and 2140, as illustrated in Figure 23. I itiigure "'2'5t-:=;—Zillus_traftes_:,i e. " rlrf the .use'r:: may .;y.ia_i.v..-we p'cé‘diciif=lf_r;i'ir;é'='i'a_"e’i_tls'ir.ifc._l§y=:rt:s:mng-;.lr:eeia'li_iisr 4:l.:D.%:‘aiie.adiiofsiiourrent .tim'e.£ loss; Figure 25 illustrates in View 2500 an interactive dialog box 2510 through which the user may define the conditions to set an alarm for a particular-risk. The user achieves this by selecting the particular risk and then setting upper and lower thresholds for the patient state probability associated with this risk. No alarm is activated as long as the patient state probability is between the upper and the lower threshold. The alarm is activated when the patient state probability crosses the threshold. Once the alarm is activated the system 100 may notify a list of chosen people, or send the notification to another clinical system. Any of module 122-124 may actually store their respective threshold data ranges and initiate the trigger depending on the specific parameter.

Figure 26 illustrates in View 2800 yet another possible visualization of the patient risk trajectory, i.e. the evolution of the patient states’ probabilities associated with particular risks.

W0 2013/142268 PCT/US2013/031512 The user may choose what time series of patient state probabilities heishe wants to display, and the system plots these probabiiities against time.

Figure 27 illustrates in view 2700 how the system 100 may directly present the probability density functions of various internal state variabies. Specificaliy, in the example, the estimated PDF of oxygen cieiivery is plotted in graph 2710 as a function of time, with darker colors corresponding to higher likelihood. Similarly, the estimated PDF of mixed venous oxygenation saturation (SvO2) is plotted in graph 2720 and compared with actual measurements (dark circles). information Sharing Among Users Figure 28 in view 2800 iiiustrates an example of a tagging definition interface 2810 that enables ciiniclans to mark specific instances of time or specific periods of time 2820 that are of interest or represent important points in the ctinicat course, is. a tag. Tags may be shared or sent via dialog box 2830 to specified recipients, or may be included in notes or any other part of the user interface. Users may be able to annotate a tag with particular comments or observations via dialog box 2840, and tags may be classified into categories from menu Hat 2850, for example, a tag may represent a change in medication dosing, an intervention, a note regarding monitors or measuring equipment, etc. Tags and their respective time~series markings may be color coded to indicate various properties, such as their category. For example, green tag-marks on a time series may represent changes in medication, red tag- marks may represent interventions, and yellow tag marks may represent periods of heightened concern. When setting a tag, the user may be prompted to define time instance or the time period, the category of the tag, the annotation for the tag, and how the tag should be handled by the system. Furthermore, annotations may be suggested by using natural language processing to convert the etiologies of the condition into note form.

Figure 29 illustrates in View 2900 a Newsfeecl View 2910 comprising tags, notes, or information taken from external sources, such as the time of a blood draw as taken from an electronic medical record (EMR). The Newsteeo 2910 may allow clinicians to View and post events, periods of interest, interventions, notes, tags, etc., which are posted by other clinicians.

Clinicians may View the entire Newsfeecl, or sort it based on Tag category, hazard level, etc.

Further, clinicians may search for tags based on keywords, intervention type, time of stay, source of information, etc. Entries 2920 ~ 2926 on Newsfeed 2010 may indicate any of the category, source of tag, and patient overview, either in words, or as a picture, such as the risk profile. 28 36 W0 2013/142258 PCT/US2013/031512 Figure 30 illustrates in View 3000 a Condition Summary View 3010 through which the clinician may request a condition summary by selecting or clicking a particular patient state. The Condition Summary View 3010 then may present clinicians with a description of a particular state, including both definitions in window 3020 of the state, and information regarding how the system arrived at the conclusion about this patient state probability. This View 3010 may provide the likelihood and hazard level of the patient state, the definition of the patient state in terms of iSV thresholds. and the likelihood of each attribute defining the state, as illustrated and can also provide a natural language description and evidence window 3030 of evidence contributing to the patient state, by translating the iSVs PDFs into a qualitative textual description, or by directly presenting numerical information regarding the evidence. As an example, Figure 30 illustrates the Condition Summary View 3010 for the patient state shock due to low cardiac output. Here, the Condition Summary View 3010 presents the probability of the "shock due to low cardiac output” state, and the hazard level of the state shown in color or dotted hatching. Further, the definitions of shock {mixed venous saturation below or equal to 45%), and low cardiac output (cardiac output below 3.2 liters per minute per meter squared), along with the probabilities that each of these are satisfied (e.g., 40% for shock and 30% for low CO) is presented. The evidence window 3030 conveys the information which leads to the assessment of “Shock due to low CO”. in this example, the system has converted the information regarding the probabilities of the etiologies into a textual form. specifically that the estimated probability is mainly driven by the fact that there is sub nominal pulse pressure (systolic blood pressure minus diastolic blood pressure) indicating reduced stroke volume.

Figure 31 further illustrates in view 3100 the ability for the user interface to include and display reference material, which may be accessed through the lnternet, or stored within the system 100 or remotely accessible thereby. Selecting the Learn More button 3110 in the Condition Summary View 3010 may bring up 3 Learn More View 3120, which shows reference information associated with the Condition Summary View 3010. Reference information may include causes, interventions, common comorhidities, anatomy, relevant publications, etc.

Furthermore, this feature may serve as a training tool to familiarize clinicians with managing the particular patient population, or treatment strategies.

HLHS Stage 1 Example _ The following description explains how the disclosed system 100 and techniques can be applied to the modeling of the clinical course of a speciﬁc patient population under intensive oare——post—operatively recovering l-lypoplastic Left Heart Syndrome patients after stage one palliation. 27 W0 2013/142268 PCT/US2013/031512 Hypoplastic Left Heart Syndrome is a congenitai heart defect, which is manifested by an underdeveloped left ventricle and left atrium. As a result, patients suffering from this condition do not have separated systemic and pulmonary biood flows, but instead the right ventricle is responsibie for pumping biooct to both the body and the tunes. Therefore, the hemodynamic optimization during intensive care invotves managing the fractions of the blood fiow that pass through the Eungs (puimonary flow Qp) and the body (systemic flow Q5). The optimai hemodynarnic state is reached when: adequate tissue oxygen delivery, D02, is achieved for a pulmonary to systemic biood flow ratio, denoted Qp/C19, of 1. Often, to reach this optimal state, the patient physiology passes through other iess beneficial states, and the correct identification of these states and the application of a proper treatment strategy for each one of them define the quaiity of the post—operative care.

:Véneb'ie' " ‘ Description H Units Type 902 indexed Oxygen Delivery mi. Ozimin/m2 Dynamic V02 indexed Oxygen Consumption mi. Ozlminlmz Dynamic PVR Puimonary Vascular Resistance mm Hg/L/min/mz Dynamic I Si/R System Vascuiar Resistance mm Hg:’L!rninim2 Dynamic APVR Change in PVR per time step b mm Hg/Limingmz Dynamic ASVR Change in SVR per time step mm Hg,ii_/mingmz Dynamic Hb Hemoglobin gldt. Dynamiciobsen/ed HR Heart Rate Beats per min Dynamic/Observed Spi/O2 Puimonary Venous Oxygen Saturation % Dynamic SaO2 Arterial Oxygen Saturation % Derivediobsenied SVO2 Systemic Venous Oxygen Saturation % Derived/Observed SD02 Puimonary Venous Oxygen Saturation % Observed r; Aortic Compliance Dynamio ABPm it/lean Arterial Biood Pressure mm Hg Derived/Observed CVP Central Venous Pressure mm Hg Dynamic/Observed LAP Left Atrial Pressure mm Hg Dynamic/Observed RAF’ Right Atrial Pressure mm Hg Dynamic/Observed AP Puise Pressure mm Hg Derivediobsewed CO Totat Cardiac Output i_/minymz Derived Qp Puimonary rate of btood fiow Llmin/mg Derived Q5 Systemic rate of blood fiow Liminimz Derived QPIQS Ratio of Q9 to Q3 Derived AQp_.QS Change in Ratio of Q9 to Qs per time Derived step 28 W0 2013/ 142268 PCT/US20I3/031512 CD02 Oxygen Delivery feedback constant 7*’ CV02 Oxygen Consumption feedback constant CH5 Hemoglobin feedback constant -- 03 Aortic compliance scaling constant ~—. 0,, Aortic compliance offset .«.~ 05 Hemcgiobin oxygen carrying capacity _ » Table ‘1 Table 1 lists state variables that may be used in the model of HLHS physiology after stage 1 palliation, the variable description, units, and type oi variable. A person reasonably skiiled in the relevant arts will recognize that though these variables encompass circulation, hemodynamic, and the oxygen exchange components of HLHS physiology. the modeis can be altered or enhanced with any additional physiologic components such as ventilation, metabolism, etc. without altering the premise of the disclosed invention.

Figure 32 depicts a general Dynamic Bayesian Network (DBN) that may be employed to capture the physiology model of the HLHS stage 1 palliation patients. The graphical model illustrated conceptually in Figure 32 captures the causal and probabilistic relationship between the variables of the model. in the DE-IN, the state variables are organized into three groups: dynamic variabies 3210, derived variables 3220, and observed variables 3230. Dynamic variables 3210 are variables whose values change over time based on a dynamic probabilistic model to be described below. Derived variables 3230 are quantities that depend on the dynamic variables with some functional relationship. These variables are computed or derived when required from the latest dynamic variables and are therefore also dynamic in nature. Observed variables 3230 are those variables that are measured directly by one of the sensors connected to the system and the patient. Observed variables 3230 represent instances of the true dynamic or derived states variabies that have been observed under noise.

Figure 33 lists several equations that may be used to model the dynamics of the HLHS stage 4; physiology. The model consists of four main types of stochastic models. The first type of model is a stochastic feedback control model (eqs. 1, 2, and Y). These variables have a riominai value that the body maintains, but are disturbed by some random process off of this nominal value. The strength at which the body attempts to maintain these values is decided by the feedback constant. The second type of model is a drift diffusion process (eqs. 3 and 4). These variables are driven over time by a random white noise process and a drift rate process. The third type of modai is a simple random walk process {eqs 5, 6, and 8). The last type of dynamic 29 WO 2013/142268 PCTI US2013/031512 mo-dei is a memory—iess process model in which the variable has no relationship to the variabie at the previous time period, but is sirnpiy a random variable which value changes at each time instance according to some predefined distribution over the proper support of the variable, i.e. a gamma distribution over the entire positive real line with parameters A and B. (eqs. 9, 10, and ‘l 1). With the exception of equations 9, 10, and ii, the driving noise for each dynamic modei is independent white Gaussian noise.

Figure 34 depicts exampie equations that may be used to abstract the relationships between the dynamic varialoies in the mode! and the derived variables. Some of these functions! reiationships are true for general human physioiogy, but many are the result of the parallei circuiation physiology that is specific to the HLHS popuiation. Equations 12-15 describe relationships for variables that are measured directiy. Equations 1648 describe functions! relationships for variabies that are of highest interest when managing care of HLHS patients post-surgery, specificaliy Cardiac Output (00) and Pulmonary to Systemic Fiow ratio (Qp:Qs).

These variables cannot be measured directly without compiex procedures.

Given these functionai relationships and the definition of the dynamic states, Figure 35 now provides a possible observation modei that may be used to relate the derived variables with . the available sensordata. Each observation model is a conditional Gaussian relationship. Under this model, the measurement received from the sensor represents a direct observation of the underlying state variable corrupted by additional independent Gaussian White noise with some variance. The figure depicts the observed quantity as the underiying state variabie with a tiide over the variabie name. in this implementation, different sensors can map to the same underlying state variabte, but with potentially different noise ieveis. For exarnpie, SpO2 as reported by a pulse oximeter measures the underlying physiology state, SaO2 or arterial oxygen saturation non—im/asiveiy. An intravenous catheter inserted directiy into the arterial blood stream aiso measures this quantity but in an invasive way. The catheter measurement shouid be a more accurate measurement than the puise oximetryyin this model. this is handled with a smaller measurement variance, R. in the HLHS physiology observer, inference over the DEN is performed using a particle filter. As described earlier, a particle filter is an example of an approximate inference scheme that uses Monte Carlo samptes of the internal state variabies to approximate the probability density function of each state variables with an ernpiricai distribution based on the number of paiticies. The fitter uses a process known as Sequentiai importance Sampling (SlS) to continuousiy resampie pariicies from the most recent approximate probability distribution. in the filter, each particle is assigned a weight. When a new observation or measurement arrives, the WO 2013/142268 PCTfUS2013/031512 weights of each particle are updated based on the likelihood of the particular particle given the observation. The particles are then resampied based on their relative updated weights, the particles with the highest weights being more likely to be resampled than those with lower weights.

Figure 36 iliustrates possibie attributes, patient states, and an etiology tree that may be used by the ciinicai trajectory interpreter module 123 in the case of the HLHS Stage 1 population. The variable total cardiac output defined as the sum of the systemic and the pulmonary blood flows is used to define low and normal total cardiac output, the Qprcls ratio is used to derive low, balanced, and high Qp:Qs ratio, the value of hemoglobin concentration Hgb is used to derive low and normal hemoglobin, and the value of the mixed venous oxygen saturation, Si/O2, is used to derive the attributes hemodynamic shock and no hemodynamic shock. This results in eight possible states deﬁned as fotiows: 1) Shock caused by low total cardiac output, as the presence of both of the attributes Shock and low total cardiac output; 2) Shock caused by low hemoglobin as the state with attributes shock, normai total cardiac output, and low hemoglobin; 3) shock from unknown causes as the state with the attributes shock, normal total cardiac output, normai hemoglobin, and balanced circulation; 4) shock cause by low Qp:Qs as the state with the attributes shock, normal total cardiac output, normal hemoglobin, and low Qp;Qs; 5) shock cause by high Qpzcts as the state with the attributes shock, normai total cardiac output, normai hernogiobin, and high Qp:Qs; 6) normal circulation as a state with the attributes of no shock and normal circulation; 7) tow Qp:Qs as the state defined by the attributes of no shock and low Qp-:Qs; and 8) high Qp:Qs as the state deﬁned by the attributes of no shock and high Qp:Qs. Figure 35 also illustrates a possible realization of an etiology tree describing the relationships between the attributes and the patient states. Using the particles approximation of internal state variables the probability of the eight states can be calculated by calculating the resistive fraction of particles within each state.

Example of Applying The Risk-Based Monitoring System in Conjunction With Evaluating Consequences Of A Possible Treatment Another possible application of the risk based monitoring system is to assist clinicians when deciding whether to appiy a particular treatment, one example being blood transfusion.

Transfusion of blood and blood products is a common in-hospital procedure. Despite that blood transfusion indications and policies are neither well established nor consistently applied within or between medical centers. Multiple studies have demonstrated variation in transfusion practices among different hospitals, practitioners, and procedures. This variation persists even when applied to a single procedure (ag. coronary artery bypass graft surgery). 31 W0 2013/142268 PCT/US2013/031512 Moreover, biood transfusion has been increasingly recognized as an independent risk factor for morbidity and mortality. Specific events and outcomes associated with transfusion include sepsis, organ ischemia, increased time on ventiiation support, increased hospital length of stay, and short- and long-term morbidity. This relationship is proportional to the transfusion volume, and evidence suggests that high hematocrit values may be detrimental.

Understandably, researchers conventionally recommend transfusion policies aimed at achieving an informed tradeoff between the risks and benefits.

Setting robust and effective transfusion policies has been proven to be a difficult task.

The consensus in the medical community is that simple policies — such as hemoglobin threshoid policies - do not provide adequate guidance. This is due to the compensatory nature of hemodynamic physiology; patients have a variable capacity to toierate low hemoglobin.

Consequently, effective transfusion decision-making must integrate factors such as compensatory reserve, intravasoular volume, hemodynamic stability, procedure type, and other patient data. Thus, there is an essential need for blood management policies that will utilize the fuli spectrum of relevant clinical variables and determine the risk/benefit ratio of transfusion.

This is exactly afforded by applying the risk based monitoring system.

Figure 37 illustrates one possible environment in which the risk based monitoring system can be applied to assist clinicians in deciding whether to apply a particuiar treatment. Consistent with the disclosure, a patient 101 is being monitored with multipte measurements 3910, both intermittently and persistently. The persistent measurements may include mixed venous oxygen saturation (83102) 3911. systolic, diastolic and mean arterial blood pressures (ABP sidim) 3913, heart rate 3916, monitored via a bedside monitor. The intermittent measurements may include blood pH 3915, hemoglobin concentration (iigb) 3912, and lactic acid concentration 3914 monitored through periodic blood works. These measurements 3910 are fed into an enhanced risk based monitoring system 3940 with treatment evaluation, which, in addition to the previously disclosed physioiogy observer module 122 and clinical trajectory interpreter module 123, consists of severai other modules. A possible treatment complications determination module 3924 receives information from the ciinicai traiectory interpreter module 123, together with information about the patient demographics 3931 and type of procedure 3932. With the information, this module BQZ4 queries an outcome database 3943 and receives back information of what the probability of different complications can be given that a) the patient is in particular patient states with particular probabiiities; b) the patient is of certain demographics (age, sex, etc); o) the patient has had a particular type of procedure; ti) and any combinations thereof of a) b) and c). On the other hand, the outcome database 3943 can be populated by 32 U1 W0 2013/142268 PCT/US2013/031512 using outcome studies 3990 derived from retrospective studies 3991, randomized cllnicai trials 3992, institution specific outcomes 3993 determined from previously collected patient data for a particular institution, and any combination thereof of the proceeding elements.

When the possible treatment compilcations determination module 3942 determines the possible complications, it feeds this information back to an enhanced visualization and user interactions module 3941. The enhanced visualization and user interactions rnoduie 3941 combines the patient—specific risk based monitoring performed by the physiology observer module 122 and the ciinicai trajectory interpreter module 122, with the evaluation of probable complication. This affords the system to provide a superior vantage point from which the ciinician 3920 can better recognize risks and benefits of treatments such as blood transfusion, and respectively more efficiently and effectiveiy decide whether to administer this treatment 3960 or not.

Figure 38 shows a non-limiting example set of patient states reievant to blood transfusion that may be used to inform the blood transfusion decision. The states contain information about the dynamics of the hernogiobin (decreesinglstable/increasing) and the hemodynamic compensation for reduced blood oxygen carrying capacity. in uncompensated patients, the hemodynamic autmregulation mechanisms become incapable of overcoming the depicted blood oxygen carrying capacity, marking the onset of anaerobic metabolism. These seven states can be determined through three internal state variables: oxygen delivery, hemoglobin, and rate of hemoglobin production/loss. Specificaliy, when hemoglobin is above 13 mgfdL, it is assumed that there is no Hgb related pathology 4001. When Hgb is lower than 13 mgldL, there are six other states, determined through five different attributes. From the oxygen cieiiirery ISV, the system can determine whether the patient is compensated or uncompensated, e.g., it may be assumed that D02 above 400 mi/minlmz, for ventiiaied and paralyzed patient, indicates compensation, and below this value uncompensated patient. The other three attributes are determined from the Hot) rate iSV and are stable 4013 and 4014 (the rate is close to zero), increasing 4011 and 4012 (the rate is positive), and decreasing 4015 and 4016 (the rate is negative).

Using The Risk Based Monitoring System With Standardized Clinical Plan Yet another appiloation of the risk based monitoring system is in applying standardized medical pians. Figure 39 illustrates one possible embodiment‘ of this application. Specificaliy, the data from the clinical traiectory interpreter module 123 is fed to a treatment query module 4142. The treatment query module 4142 queries a treatment plan database 4143 based on the determined patient risks. The treatment pier: database 4143 specifies a map between patient 33 WO 2013/142268 PCT/US2013/031512 risks and treatments. When the database 4143 returns a treatment pi:-an, it is represented to the ciinioian 3920 by an enhanced visualization and user interactions moduie 41-41 with plan. The clinician 3920 can then make ciinicai decision 4190 with respect to patient 101. The user decision, the context under which it was taken, (the calcuiated patient risks, the estimated iSVs, and other possible patient data at the time of the decision) are then recorded to a decision data base 4144. The decision database 4144 then can be compared to patient outcomes and utilized in the improvement of the treatment pian.

Figure 40 iliustrates an exampie appiication of the risk based monitoring system 4240 combined with a specific type of standardized ciinicai pian. The particuiar example considers the medical decision whether to treat the patient with nitric oxide; Nitric oxide is -a pulmonary vesodiiator and is used to treat high puirnonary vascular resistance and ensuing pulmonary hypertension, which can cause reduced cardiac output. in the exampie, the medical plan uses the risks caiculated by the ctinicai trajectory interpreter moduie 123 and stratifies 4230 them into two categories: low risk and high risk. if the risks are low 4201 the recommended decision is not to treat 4202, respectively, if the patient is classified as being in high risk the recommended decision is to treat 4203. The provider can then make a decision to either folicw the recommendations 4250 or disregard them 4260. if the provider chooses to disregard the treatment recommendation for a high risk patient, he needs to provide justification 4220.

Likewise, if the provider chooses to treat a low risk patient, he also needs to providejustitication 4210. Justifications 4210 and 4220 in conjunction with patient outcomes may be utilized to refine the risk stratification 4230 and risk—based monitoring system 100.

Figure 41 iilustrates the example risk stratification that may be employed by the system in the context of Nitric Oxide treatment. Specificaliy, it assumes that the patient can be in four different states: State 1: Low CO, Norma! PVR; State 2: Low 00, High PVR; State 3: Normal CO, High PVR; State 4: Normal C0 Norma! PVR. A patient being in tow risk may be defined as P(State 1) < 10% and P(Stete 1) + P(State 2) < 30%. Sirniieriy high risk may be defined as: P(State 1) > 10% and P(State 1) + P(State 2) > 30%.

Using The Clinical Risk Assessment System in Outpatient Care Of Chronic Conditions Yet another embodiment of the present disclosure aiicws the ciinical trajectory tracking in outpatient care. Outpatient care of chronic conditions invoives sporadic patient assessment from intermittent visits, patient selfevaiuations, and observations from caregivers. This ieads to uncertainties in determining the patient ciinicai course and the efficiency of the prescribed treatment strategy. To achieve effective patient care management, clinicians must understand and reduce these uncertainties. They have two main decisions at their disposal: 1) scheduie 34 WO 2013/142268 PCT/US2013/031512 visits, prescribe tests, or solicit self-evaluation (or caregiver evaluations) to improve their understanding of the clinical trajectory; and/or 2) prescribe changes of medication or medication closing to achieve a better tracle—off between the likelihood of improvement and possible side- eifects. To inform this decision making process, there is a need for processing the available patient information in a way that conveys the clinical trajectory, the uncertainty in its estimation, and the expected effect that different treatment strategies may have on the future evolution of the clinical trajectory.

As a non-limiting example embodiment of the risk based monitoring system to outpatient clinical trajectory tracking, we consider its application to the outpatient care of Attention Deficit and Hyperactivity Disorder (ADHD) of pediatric patients. Figure 42 illustrates possible patient states that may describe the clinical trajectory of an ADHD patient. They are the same as the ones used by the Clinical global impression-improvement scale: 1) very much worse; 2) much worse; 3) worse; 4) No change; 5) Minimally improved; 6) Much improved; 7) Very much improved. The Patient State Distribution (PS0) is the set of probabilities that the patient is in any of the seven states, given all available information and observations.

To evaluate the patient state, a clinician may either schedule an office visit for direct examination, or may request at Vanderbilt diagnostic test from family members or teachers (the test is modified depending on the respondent, teacher or parent). Figure 43 lists the available patient evaluation modalities as M1, M2 and M3. Models may be used to map the test questions and answers into the states of the patient. Both the clinical evaluation and the test—based evaluation are associated with uncertainty that prohibits the exact determination in which of the seven states the patient currently resides.

The dynamic model or the patient evolution from state to state may be abstracted by a Dynamic Bayesian Network (DEN) as the one shown in Figure 44. in Figure 44, the arcs’ directions signify statistical dependence, i.e., the connection from “Patient state @ it" 4601 to "ll/ll”, signifies the probability density function (PDF): P(M1iPatl'ent state @ £1). Similarly, the depicted DEN illustrates that “Patient state @ t2" 4802 (the patient state at a particular time t2), is conditioned on the “Patient state @ it” (the patient state at the previous time increment ii). in the spirit of the present disclosure, this model enables the estimation of the patient state distribution even in the absence of some or all possible measurements, eg., as illustrated in the figure at time instance t2 when M1 is missing, and at time instance t3 ("Patient state @ E3" 4603), when all measurements are missing.

Figure 45 illustrates an alternative embodiment for two predictions of how the patient state can transition in a single month given medication change or a dosage change. This '15 W0 2013/142268 PCT/US2tJI3/031512 prediction is performed based on a statisticai model derived in the following fashion: Step 1: isolate a group of patients from retrospective data which at some point of their treatment have passed through State A and received a change of treatment (ll/Jedi Dose‘! -> Med 2 Dose 2); Step 2: For each patient, set the time instance that this perticuiar event occurred to to; 3) step 3: for each patient identify what is the patient state at time to + 1 Month (1 M) (or any desired time step unit). 4) calculate the fraction of patients that transition State A ~> State i where i stands for at! seven possible patient states; 5) set the fraction as the probabiiities for transition under the particular treatment change.

Figure 46 shows one possibie embodiment and scenario of visuaiization displaying the patient clinical trajectory and risks. The user interface denotes that the patient is at “no change” state, and that this has been established by three separate measurements: office visit, teacher based Vanderbiit diagnosis, and parent based Vanderbiit diagnosis. The solid tine on the screen signifies that a medication has been prescribed to the patient (Medication 1) at week 1 of the treatment.

Figure 47 shows an evaluation of the patient and the patient trajectory at west: 9 at which point the ciinical risk assessment system determines a probability density function for the state of the patient for each of the past six weeks. The availabie measurements at this point are teacher and parent Vanderbilt diagnosis. in the iilustrated example, due to a high probability of a deteriorating patient state, the ciinician prescribes a change of medication dosing, which is depicted by a dashed red line. Additionaliy, the user inteztace shows the side effect reported by the patient—headache.

Figure 48 shows a fo|iow—up evaluation based on teacher and parent Vanderbilt diagnosis. in the example, the clinician decides a medication change depicted by a holiow tine. ﬁgure 49 shows consequent evaiuation based on art avaiiable measurements—0tfice visit, parent and teacher evaiuation, which estabiishes high probability for significant improvement.

Figure 50 shows yet another ioiiow-up at which point it is estabtished that the patient is most probably stably improved, and has been stabiy improved between the two evaluations.

Note that due to the apptied inference, the PDF for the patient traiectory is contlnuousiy estimated. However. the precision (the concentration of the PDF) is higher in the presence of a measurement.

Figure 51 shows a foliow—up evaluation of the patient and the patient traiectory in the absence of measurements. Due to the tack of recent observations, the uncertainty is increasing. 36 WO 2013/142268 PCT/US2013/031512 Figure 52 shows the state of this uncertainty given a full patient evaluation (ail measurement modalities). The inference engine propagates this uncertainty back in time to produce a more precise estimation of the patient trajectory, which helps the clinician to deduct that the patient is stable.

Figure 53 illustrates yet another possible visualization from the described system output. it shows possibie patient state transitions under changes of treatment pian, e.g., change of medication. it also conveys what possible side-effects can be expected. For every side effect. there are three stages of manifestation - mild, moderate, and severe represented with the three boxes next to each side effect in the figure. The coloring corresponds to the probability of a particular severity manifestation for each particular side-effect, with darker colors indicating higher probability.

Various examples and embodiments consistent with the present disciosure have be described in detailed above. it is to be understood that these examples and embodiments of the present disclosure are provided for exempiary and iilustrative purposes oniy. Various modifications and changes may be made to the disclosed embodiments by persons skilied in the art without departing from the scope of the present disclosure as defined in the appended claims.

What is claimed is: 37

Claims

CLAIMS:

1. A risk-based monitoring system for transforming measured data, including heart rate and blood oxygen, of a critical care patient into hidden internal state data that is not directly, non-invasively measurable with sensors, the hidden internal state being monitored by the system, the system comprising: a processor; a data reception module, the data reception module having a set of inputs for a plurality of sensors including at least a heart rate sensor and an oximetry sensor, the plurality of sensors being couplable to the critical care patient, wherein the plurality of sensors provide, to the data reception module, data associated with a corresponding plurality of internal state variables m , s=1, . . . , n, over a series of time steps t , k=0, s k 1, . . . Z, each internal state variable m characterizing a parameter physiologically s relevant to at least one of a treatment and a condition of the patient; a physiology observer module, in communication with the data reception module, the physiology observer module configured to update, via a first and second computer processes performed by the processor, the data provided by the sensors to the data reception module, wherein: the first computer process generates a conditional likelihood kernel for the internal state variables m at time t , the conditional likelihood kernel comprising a set S k of probability density functions, each such probability density function being for a distinct internal state variable m at time t ; and S k the second computer process generates, using Bayes theorem, posterior predicted conditional probability density functions for each of the internal state variables m for the time step t given the conditional likelihood kernel for the internal s k state variables at time t and a predicted conditional probability density function of k each of the internal state variables for time t ; and k a clinical trajectory interpreter module, in communication with the physiology observer module, configured to determine, based on the generated posterior predicted conditional probability density functions of the internal state variables m at time step s t , a set of possible states of a hidden internal state variable to generate a probability k 38 value representing the likelihood that the patient’s oxygen delivery, which cannot be directly, non-invasively measured, is inadequate.

2. The system of claim 1, wherein the oximetry sensor is an SpO sensor. 2

3. The system of claim 1 further comprising: a display device operably coupled to the processor; and a user interaction module configured to display, on the display device, a plurality of graphical indicators, each of the plurality of graphical indicators corresponding to one of the states of the set of possible states of the hidden internal state variable, each of the plurality of graphical indicators graphically identifying the probability that the hidden internal state variable is in a corresponding state at a given point in a range of time, the plurality of graphical indicators configured to indicate a hazard level.

4. The system of claim 3, wherein the user interaction module is further configured to display, on the display device, a clinical trajectory of the patient based on the conditional probability density function of the hidden internal state variable over at least some of the series of time steps.

5. The system of claim 1, wherein the physiology observer module compares a newly received measurement associated with an internal state variable m at time t s k with a predetermined predicted likelihood of probable measurements given previously received measurements, and does not generate a conditional probability density function for the associated internal state variable m , if the newly received s measurement is not within the predetermined predicted likelihood of probable measurements for the associated internal state variable m . s

6. The system of claim 3, further comprising a timeline controller configured to allow a user to dynamically select a plurality of points in time over the series of time steps t , the graphical indicators changing dynamically in response to a specification k 39 by the user of one of the plurality of points in time, to display the evolution of the hidden internal state variable.

7. The system of claim 1, wherein the physiology observer module generates the conditional likelihood kernel for the internal state variables m at time t , using a S k predicted conditional probability density function, generated at a previous time step, for each of the internal state variables.

8. The system of claim 1, wherein the physiology observer module is further configured to generate, at time step t , a predicted conditional probability density k function for each of the internal state variables m for time t based on the posterior S k+1 predicted conditional probability density functions for each of the internal state variable mS for the time step tk.

9. The system of claim 8, wherein the conditional probability density functions for the time step t are generated using the posterior conditional probability density k+1 functions for each of the internal state variables m from a preceding time step t , s k using the formula: P(ISVs(t )│M(t )) = ∫ P(ISVs(t ) │ISVs(t )) P(ISVs(t ) │M(t )) dISVs. k+1 k ISVsϵISV k+1 k k k

10. A method of transforming measured data, including heart rate and blood oxygen, of a critical care patient into hidden internal state data which is not directly measurable with sensors, the method comprising: providing a plurality of sensors including at least a heart rate sensor and an oximetry sensor, the plurality of sensors being configured to be physically attachable with the critical care patient; attaching the plurality of sensors to the patient; substantially continuously acquiring, by a computer, from the plurality of sensors connected with the patient, physiological data associated with a corresponding plurality of internal state variables m , S=1, . . . .n, over a series of time steps t , K=0, s k 1, . . . Z; 40 generating, by the computer, a conditional likelihood kernel for the internal state variables m at time t , the conditional likelihood kernel comprising a set of s k probability density functions of the internal state variables m for the time step t , each s k of the internal state variables describing a parameter physiologically relevant to at least one of a treatment and a condition of said patient at time step t ; k generating, with the computer and using Bayes theorem, posterior predicted conditional probability density functions for the plurality of the internal state variables m for the time step t given the conditional likelihood kernel for the internal state s k variables m at time t and predicted probability conditional density functions of each s k of the internal state variables m for time step t ; and s k identifying, with the computer, from the generated posterior predicted conditional probability density functions of the internal state variables m at time step s tk, a set of possible states of a hidden internal state variable associated with oxygen delivery (DO ), and generating a probability value representing the likelihood that the 2 patient’s oxygen delivery (DO ) is inadequate, which cannot be directly measured. 2

11. The method of claim 10, wherein the oximetry sensor is an SpO sensor. 2

12. The method of claim 10 further comprising substantially continuously displaying a clinical trajectory of the patient on a graphical user interface, the user interface being configured to display the probability of the hidden internal state variable as function of a plurality of time steps.

13. The method of claim 10, wherein the probability value associated with the hidden internal state variable indicates risk that the patient’s oxygen delivery is inadequate and the patient may be in a state of shock.

14. The method of claim 10, wherein generating the conditional likelihood kernel for the internal state variables m at time t , includes using a predicted conditional S k probability density function, generated at a previous time step, for each of the internal state variables. 41

15. The method of claim 14, further comprising generating, at time step t , a k predicted conditional probability density function for each of the internal state variables m for time t based on the posterior predicted probability density S k+1 functions for each of the internal state variable m for the time step t . S k For the Applicants REINHOLD COHN AND PARTNERS By: 42