WO2019215742A1

WO2019215742A1 - System and method for hypotensive episode prediction

Info

Publication number: WO2019215742A1
Application number: PCT/IL2019/050526
Authority: WO
Inventors: Mark Last; Daniela BEHNAM; Elad Tsur; Victor F. GARCIA; Raphael Udassin
Original assignee: B.G. Negev Technologies And Applications Ltd., At Ben-Gurion University; Children's Hospital Medical Center; Hadasit Medical Research Services And Development Ltd.
Priority date: 2018-05-09
Filing date: 2019-05-08
Publication date: 2019-11-14

Abstract

The invention relates to a method for predicting a hypotensive episode, comprising: (a) defining durations of: a sliding target window, a sliding gap window located prior to the target window; (b) a sliding observation window which is divided to several sub-windows and located prior to said gap window; and (c) defining a set of features; during a training stage: having the above definitions, preparing a model based on a retrospective database containing a mass of patients' monitoring records; and during a real-time stage: (d) continuously applying the divided observation window (e) periodically determining a set of real- time feature values for each sub-window of the divided observation window; (f periodically submitting each of said determined real-time feature values to the model and calculating by said model a probability for a hypotensive outcome episode to occur within the target window.

Description

SYSTEM AND METHOD FOR HYPOTENSIVE EPISODE PREDICTION

FIELD OF THE INVENTION

The field of the invention relates in general to medical devices. More specifically, to a system and method for monitoring and predicting a hypotensive episode in a patient.

BACKGROUND OF THE INVENTION

Hypotension, defined as clinically significant low blood pressure, is a significant risk factor to patients and requires prompt therapeutic intervention to avoid organ failure and death. In critical and the emergency care settings, unrecognized and uncorrected hypotension is one of the most significant risk factors for mortality and multiorgan failure hospital acute care settings. A hypotensive episode (hereinafter also referred to as "HE"), which originates from an internal bleeding cannot be visually observed, and a relatively long period may pass until the hypotension is suspected, measured and determined as such, leading to preventable errors in clinical decision making, treatment, and increased potential for loss of life and organ failure. The prediction of a hypotensive episode, before severe deterioration, is a challenging task for several reasons. First, the amount of time, which can be allocated by the clinical staff per patient is limited. Second, the amount of accumulated physiologic data per patient is massive in terms of both data variety (multi-channel waveforms, laboratory results, medication records, nursing notes, etc.) and data volume (length of waveform time series) . Even with sufficient time, resources, and data, it is very hard to induce an accurate estimation with respect to the likelihood of clinical deterioration with a bare-eye analysis alone. Furthermore, recent research in human decision making suggest that clinical predictions based on the subjective impressions of even highly trained professionals were less accurate than statistical predictions made by algorithms. Sixty percent of algorithms had better accuracy than clinical predictions. Daniel Kahneman concludes that algorithms are superior to experts in part because humans are incorrigibly inconsistent in making summary judgements of complex information as is characteristic in high stress high acuity environments as seen in the emergency and operating theater and intensive care units. Kahneman and others have revealed a long list of biases of judgement and choice especially overconfidence among seasoned clinicians and availability heuristic. The pediatric patient presents a unique conundrum in that despite losing up to 50% of his/her blood volume the child's will typically register a 'normal' blood pressure on conventional patient monitors. Consequently, the existing patient monitoring systems installed in intensive care units do not have the HE prediction capabilities.

Numerous trials have attempted to predict a hypotensive episode in a sufficiently advanced time, by means of analyzing continuously monitored physiologic data. The vital signs (multi-parameter temporal vital data) typically inform on the underlying dynamics of organs and cardiovascular system functioning. Particularly, it has been found that vital signs may contain subtle patterns that can point to an impending instability. While the prior art for predicting hypotensive- episodes have shown some success, there is still a necessity for improvement to make the product clinically relevant.

With respect to the clinical deterioration prediction problem, the previous works vary in several aspects. The first aspect is how an episode is defined. For example, the definition may be based on the recorded clinical treatment and on the behavior of vital signs within a specific duration of time. The second aspect is how long before an episode one aims to predict it. This period is referred to herein as a "Gap Window" (GW) or shortly "gap". Additional aspects include the extraction of future episode predictive features and the prediction algorithms used.

The 10th annual PhysioNet /Computers in Cardiology Challenge (see Moody, G. B., Lehman, L. W. H. : Predicting acute hypotensive episodes: The 10^th annual physionet /computers in cardiology challenge, Comp. Card., pp . 541-544, IEEE (2009)) has conducted a competition for studying an Acute Hypotensive Episode (AHE) . They defined AHE as an interval in which at least 90% of the time the Mean Arterial blood Pressure (MAP) is under 60 mmHg during any 30-minute window within this interval. The goal was to predict whether an AHE will begin in the next 60 min.

A generalization of clinical deterioration prediction was done in Forkan, A. R. M. , et al.: ViSiBiD: "A learning model for early discovery and real-time prediction of severe clinical events using vital signs as big data", Computer Networks 113, 244-257 (2017), where an episode was defined as one of seven critical patient conditions (such as, Tachycardia, Hypertension, etc.) which are all defined by some of four vital signs. The critical condition episode definition was at least 30 minutes in which all four vital signs deviate from their normal range. The tested gaps were 60, 90, and 120 minutes.

In most works, the future episode predictive features have used a sliding Observation Window (OW) over a record, which is a collection of vital-sign time series of one patient in a single ICU admission. Furthermore, a minute-by-minute vital- signs time series, like blood pressure and Heart Rate (HR) were usually used to extract features, while a few works also used the clinical information (age and temporal medications data) as well. Moreover, a difference between prior art works lies in the type of features extracted from the vital signs, also called parameters. Most works have used the MIMIC II benchmark database - a multi-parameter ICU waveforms and clinical database. The MIMIC II Waveform Database Matched Subset (Physionet Database) can be found at:

https : //physionet . org/physiobank/database/mimic2wdb/matched/

Some studies started by adding new knowledge-based parameters calculated based on the biological and clinical nature of vital signs, as well as their inter-relations.

Statistical features were used in the art as a source for the extraction of predictive features (also called patterns), from intervals like OWs. In, Forkan, A. R. M., et al . : ViSiBiD: "A learning model for early discovery and real-time prediction of severe clinical events using vital signs as big data", Computer Networks 113, 244-257 (2017), the extremes, moments, percentiles and inter-percentile ranges were calculated for every vital-sign, whereas in Lee, J., Mark, R. G.: "An investigation of patterns in hemodynamic data indicative of impending hypotension in intensive care", Biomed. Eng. Online 9(1), 62 (2010) interquartile ranges and slope were added. In a more pragmatic statistical approach, Chen, X., et al . : "Forecasting acute hypotensive episodes in intensive care patients based on a peripheral arterial blood pressure waveform", Comp. Cardio., 2009, pp . 545-548. IEEE (2009), several episode predictive indices were used, derived from the blood pressure signals only. These indices were six types of averages from SBP, DBP and MAP, each taken as a single feature .

The prior art has suggested the use of a sliding window having a duration ranging between 30 to 60 minutes. The sliding window was analyzed as a single statistical unit. For example, in Lee, J., Mark, R.G.: An investigation of patterns in hemodynamic data indicative of impending hypotension in intensive care. Biomed. Eng. Online 9(1), 62 (2010) the highest prediction rate of a hypotensive episode that was obtained was about AUC = 91.8%.

It is therefore an object of the present invention to provide an improved method for predicting a hypotensive episode.

It is still another object of the present invention to extract additional components of physiologic data that can enhance the forecasting of an impending hypotensive episode. Other objects and advantages of the invention will become apparent as the description proceeds.

Summary of the Invention

The invention relates to a method for predicting a hypotensive episode at a patient, comprising: (a) defining a duration of a sliding target window, and a duration of a sliding gap window which is located in time prior to the target window; (b) defining a duration of a sliding observation window which is located in time prior to said gap window, said observation window being divided to several sub-windows, thereby to form a divided observation window; and (c) providing a retrospective database containing a mass of patients' monitoring records that allow retrospective determination of an outcome of occurrence or non-occurrence of a hypotensive episode in each target window; and (d) defining a set of features ; during a training stage: (a) selecting observation windows, each being labelled as either normal or hypotensive episode outcome; (b) applying said divided observation window to each of said selected observation windows and finding a set of respective feature values for each sub-window of the divided observation window; (c) applying a classification algorithm on all the respective sets of sub-window feature values to induce a probability estimation model, while considering, among others, the labelled outcome; (d) storing the probability estimation model; and during a real-time stage: (e) continuously updating the divided observation window with new real-time patient's monitored data; (f) periodically determining a set of real time feature values for each sub-window of the divided observation window; (g) periodically submitting each of said determined real-time feature values to the probability estimation model and calculating by said probability estimation model a probability for a hypotensive outcome episode to occur within the target window.

In an embodiment of the invention, the method further comprising allowing a user to tune a probability threshold, above which an alert is issued.

In an embodiment of the invention, the determination of the probability estimation model involves consideration of inter sub-windows feature values, and wherein said probability calculation by the probability estimation model involves consideration of inter sub-windows feature values.

In an embodiment of the invention, said features are selected from: statistical features, wavelet features, cross correlation features, dynamic features, and distribution features .

Brief Description of the Drawings

In the drawings:

- Fig. 1 shows an exemplary temporal record of a patient; - Fig. 2 shows in a block diagram form a general structure of a patient-monitor for determining a hypotensive episode, according to an embodiment of the present invention;

- Fig. 3 illustrates in a flow diagram form a training stage for the preparation of the monitor's probability estimation model, according to an embodiment of the present invention;

- Fig. 4 illustrates an observation window, which is divided into four sub-windows; and

- Fig. 5 generally illustrates in a block diagram form more details of the real-time stage of a method according to an embodiment of the present invention.

Detailed Description of Preferred Embodiments

Fig. 1 shows an exemplary snapshot of a temporal admission record. Typically, while during hospitalization, a monitor continuously acquires a variety of vital signals that reflect the patient's condition during his entire stay - in this example, signals relating to: (a) A patient's Systolic Blood Pressure (SBP), (b) The patient's Diastolic Blood Pressure (DBP), (c) The patient's a Mean Arterial blood Pressure (MAP), and (d) The patient's Heart Rate. The upper graph 10b shows an expansion of a portion of the lower graph 10a. The present invention reliably estimates the probability of a hypotensive episode occurrence within the future Target Window (TW) . The probability estimates are provided at a pre-defined frequency (e.g., every 1 minute) at the end of a sliding Observation Window (OW) , i.e., the start of a sliding Gap Window (GW), which therefore defines a minimal period for alert, prior to the occurrence of a hypotensive episode. For example, the duration of the Target Window may be in a range between 30-60 minutes, the duration of the Gap window may be in a range between 60-120 minutes, and the duration of the Observation Window may be in a range of between 30-240 minutes. The above durations are given as an example only, and may vary.

Fig. 2 shows in a block diagram form a general structure of a patient-monitor 10 for estimating the probability of a hypotensive episode, according to an embodiment of the present invention. The operation of the monitor 10 has two stages: (a) an initial training stage, which is typically performed once at one or several hospitals, and before the actual operation, and (b) a real-time monitoring stage, which continuously takes place during the patient's stay at the ICU. As will be discussed in more details hereinafter, the monitor

10 contains, among others, a probability estimation model 13 which is induced during the training stage.

Fig. 3 illustrates in a flow diagram form the training stage 110 for the construction of the monitor's probability estimation model 13 (shown in Fig. 2), according to an embodiment of the present invention. In step 112, a retrospective database, which contains a sizable collection of, for example, at least several thousand patients' monitoring records is provided. For example, the MIMIC II database (or similar) may be used for this purpose. The MIMIC

11 contains in digital form a mass number of historic signal- records (such as SBP, DBP, MAP, and HR) of patients, which allow to determine retrospectively for each signal record if and when a hypotensive episode has occurred. In step 114, a divided observation window (hereinafter, DOW) , a gap window (GW) , and a target window (TW) , particularly their durations, are defined. The DOW is a sliding observation window, which is divided into several sub-windows. As will be discussed hereinafter, a separate set of features is calculated for each of the sub-windows of the DOW. In step 116, a set of features is defined, to be calculated separately for each of the sub windows. Fig. 4 illustrates an observation window 100 having duration of To, (for example 120 minutes), which is divided into four sub-windows ( Subwi-Subw4 ) . A separate set of features SubFi - SubF₄ is calculated for each of the sub-windows, respectively. The duration of the DOW may vary between, for example, in a range of 30-120 minutes. The number of sub windows may also vary.

Back to Fig. 3, in step 118 a group of as many as possible observation windows of different patients, each of which can be labeled with either a normal or a hypotensive episode (HE) outcome at their TW, are selected for analysis from the available signal records. In step 120, the divided observation window (DOW) is applied to each said observation window to extract the respective values for each of the sets of features SubFi to SubF₄. In step 122, a classification algorithm is applied on all the sets of respective feature values of SubFi to SubF₄ of the selected observation windows to induce an outcome probability estimation model (13 in Fig. 2) .

A divided observation window (DOW) of a same structure is used during both the training and the real-time phases of the invention. A variety of extracted features may be used by the invention, such as: mean, median, standard deviation (Std) , variance, in-terquartile range (Iqr), skewness, kurtosis and linear regression slope that are calculated for every time series in each of the sub-windows. As will be discussed in more details hereinafter, the invention preferably also uses cross-correlated features 130 (shown in Fig. 4) and wavelet features. The entire observation window, as well as its plurality of sub-windows, are in fact sliding windows that are updated as the time progresses, for example, every minute. The update is performed with respect to new signal data which is added (at the "front" of the observation window) , and "old" data which is removed (at the "rear" of the observation window) , and with respect to the periodic calculations that are made to determine updated values for the features.

Back to Fig. 2, during the real time monitoring phase, real time signals Si-S_n are provided to a signal receiving unit 11, and conveyed 20 to the features-extraction unit 12. It should be noted that the terms "units" referred to herein are used for convenience of description only, and it should not be interpreted that they are necessarily physical units. The features extraction unit 12 periodically calculates in real time feature values 14 for each of the signals 20, and forwards these calculated feature values to the probability estimation model 13 (that was previously induced during the training stage) . Based on the set of features values 14 at its input, the probability estimation model 13 calculates an outcome probability 17 for an occurrence of a hypotensive episode during the future target window. The outcome probability 17 if forwarded to a threshold unit 14. A user can tune a threshold 18 for issuing an alert 19, based on a given probability level 17 at the input of unit 14, thereby to reduce erroneous alerts.

Fig. 5 generally illustrates in a block diagram form more details of the real-time stage of the method of the present invention. As noted, the observation window is divided into n portions, forming a sliding divided observation window DOW. The divided window virtually "slides" from the left to the right (in the direction shown by arrow 307) as the time progresses. A gap window GW separates between the "front" 302 of the DOW and the "rear" 304 of the target window TW (the "front" of a window or sub-window defines an instant of time later compared to the "rear" of a same window) . During a real time operation, the "front" sub-window, in this case SubW_n, is updated continuously with new data (reflecting one or more time series), while older data (that was entered earlier in time) at the "rear" 306 of SubW_n moves to the previous window SubW_n-i (SubW₃ in Fig. 5) . The same procedure is repeated with respect to all the sub-windows of the DOW. Every period of, for example, 1 minute, an updated calculation for the features SubFi - SubF_n is performed separately with respect to the current time series in each respective sub-window. All the calculated features are forwarded to the probability estimation model 313 substantially simultaneously. Having the calculated features SubFi - SubF_n, the probability estimation model 313 calculates the probability 317 for a hypotensive episode to occur within the future target window. A user of the monitoring unit may adjust 318 a threshold 314 for the issuance of an alert, based on the level of probability 317. As shown, in contrast to the prior art, the present invention uses a divided observation window, and a set of features that are separately calculated for each of the sub-windows. As a result, the invention in fact comprises many more features for either a separate analysis or for an inter sub-windows analysis by the probability estimation model 313. As will be discussed in more details hereinafter, the inventors have found that the rate of success-prediction is significantly increased by about 2% (i.e., from about 91.8% to 93.7%) compared to the highest prediction rate by prior art methods. Moreover, while the probability estimation model was induced based on the MIMIC II database, the above real-time prediction rate was obtained based on patient's data from two other hospitals, namely Soroka in Beer Sheva, Israel, and Hadassa in Jerusalem, Israel, and the improvement has consistently appeared .

Further Discussion and Experiments

Several experiments were performed to verify the concept of the invention. The experiments predicted a patient condition (hypotensive or normotensive) within a 60-min gap. A hypotensive episode was defined as a 30-min target window in which at least 90% of its MAP values are below 60 mmHg . Any valid target not meeting this criterion was labelled as normotensive. At the prediction time, each sub-window set was labelled with respect to its corresponding target. Considering the implementation of the method in the clinical setting, two alternative prediction modes were distinguish: (i) all-time prediction, where the assumption is that episode prediction is needed continuously throughout the stay of a patient in the ICU, regardless of the clinical condition at the prediction time; (ii) exclusive prediction, where episode prediction is needed only when the patient is not in a currently recognized hypotensive episode (the last 30 minutes of the ICU stay are not an hypotensive by definition) . The experiment assumed, for example, that a set of a mean MAP values from four, three, two and one hours before a same target window will be more informative for predicting the hypotensive episode than the mean MAP value of a single 4- hour observation window.

Three basic vital signs were used to derive two additional parameters for each record: Pulse Pressure calculated by PP = SBP - DBP, and Relative Cardiac Output calculated by CO = HR X PP . Next, three group of features were extracted from each sub-window as detailed below.

Statistical features: mean, median, standard deviation (Std) , variance, in-terquartile range (Iqr), skewness, kurtosis and linear regression slope are calculated for each of the 6 parameters (48 features) . Missing values are ignored.

Wavelet features: multi-level discrete decomposition of each vital sign can be conducted with DB wavelets. The decomposition of a single time series (signal) X is denoted by WX = [a_n dn d_n-i · · · di ] , where n is the decomposition level (window size depended) , a_n is the signal approximation, and dk is the detail signal of level k. The elements in W_x are then utilized as features by calculating the relative energy for each of them (a total of 24-42 features) . Missing values are interpolated over the record time series. Cross-correlation features: the cross correlation of two time series X = {x_lf x_2r x_n} and Y = {yi, y₂, ..., y_n} were defined by pXY = ni åXiyi and calculated for each pair of vital signs (15 features overall) .

Characteristics of sample distribution:

(a) Energy: absolute energy of the time series which is

(b) Count above/below mean: the number of values in x that are higher/lower than the mean of X.

(c) Longest strike above/below mean: the length of the longest consecutive subsequence in x that is bigger/smaller than the mean of X.

(d) First/last location of maximum/minimum: the first/last location of the maximum/minimum value of x.

(e) Approximate entropy (ApEn) : quantifies the amount of regularity and the unpredictability of fluctuations over time- series data. A low value of the entropy indicates that the time series is deterministic while a high value means that the time series is unpredictable (randomness) .

Features derived from observed dynamics :

(a) Autoregressive (AR) coefficients: the autoregressive model specifies that the output variable depends linearly on its own previous values and on a stochastic term. The unconditional maximum likelihood of an autoregressive AR(k) process is fitted. The k parameter is the maximum lag of the process, it is set to 10 to consider the last ten observations: (b) Discrete Fourier Transform (DFT) Coefficients: computed by Fast Fourier Transform (FFT) algorithm. The resulting coefficients are complex, therefore it is split to four parts: the real part, the imaginary part, the absolute value and the angle in degrees.

(c) Continuous wavelet transform-coefficients: calculates a Continuous wavelet transform for the Ricker wavelet, also known as the Mexican hat wavelet.

(d) Autocorrelation: is the correlation of a signal with a delayed copy of itself as a function of delay.

(e) Mean change: the mean over the absolute differences between subsequent time series values which is

+ 1— X_j .

(f) Sum values - the sum over the time series values:

(g) Absolute sum of changes: the sum over the absolute value of consecutive changes in the series x.

+ 1— c_;| .

(h) Number of peaks: calculated by the number of peaks of at least support n in the time series x. A peak of support n is defined as a subsequence of x where a value occurs, which is bigger than its n neighbors to the left and to the right.

(i) Linear regression: linear least-squares regression for the values of the time series versus the sequence. Slope and intercept of the regression line, the correlation coefficient were calculate, the two-sided p-value for a hypothesis test whose null hypothesis is that the slope is zero, using t-test. The total amount of features extracted from a sub-window set is equal to the number of sub-windows N multiplied by the feature set size. If, for example, the observation window is a 60 minutes window, a and the observation window is divided to 4 sub-windows, the window-set configuration may be 4 ^c 98 = 392.

Each instance in the training dataset is composed of a sub window set feature vector and a class label, which is positive or negative (the target is either hypotensive or normotensive, respectively) . Before training a binary classifier, the training dataset was both normalized (to zero mean and unit standard deviation) and under-sampled to overcome a possible imbalance. The Extreme Gradient Boosting (XG-Boost) classifier, was used for the probability estimation model. XG- Boost is a scalable implementation of the Gradient Boosting ensemble method that includes some additional abilities, like feature sampling (in addition to instance sampling) for each tree in the ensemble, making it even more robust to feature dimensionality and helping to avoid overfitting. Moreover, considering this study minimum training dataset size of approximately 2.1k instances together with the maximal feature vector size of 392 features, the built-in feature selection capability of XG-Boost was found to be important.

The classifier produced a posterior probability of the positive class, which could lead to a hypotensive episode alert depending on the probability threshold determined from a Receiver Operating Characteristic (ROC) curve.

The results of cross-dataset experiments for an all-time prediction mode have shown a relatively small drop in improvement when the training was performed in one dataset (in this case a dataset from MIMIC II), while the testing was performed in another (0.1-0.5% in AUC) . Nevertheless, it has been found that the use of the divided observation window outperforms other methods that use a single window in terms of the AUC metric, even when applying the probability estimation model to a new dataset. The experiment has also shown an average increase of 2.5% in valid prediction times when using the divided observation window in comparison with a single OW in the size of N x SubW as a result of replacing some invalid windows with their valid sub-windows.

While some of the embodiments of the invention have been described by way of illustration, it will be apparent that the invention can be carried into practice with many modifications, variations and adaptations, and with the use of numerous equivalents or alternative solutions that are within the scope of a person skilled in the art, without departing from the spirit of the invention, or the scope of the claims .

Claims

1. A method for predicting a hypotensive episode at a patient, comprising :

- defining a duration of a sliding target window, and a duration of a sliding gap window which is located in time prior to the target window;

- defining a duration of a sliding observation window which is located in time prior to said gap window, said observation window being divided to several sub-windows, thereby to form a divided observation window; and

- providing a retrospective database containing a mass of patients' monitoring records that allow retrospective determination of an outcome of occurrence or non-occurrence of a hypotensive episode in each target window;

- defining a set of features; during a training stage:

- selecting observation windows, each being labelled as either normal or hypotensive episode outcome;

- applying said divided observation window to each of said selected observation windows and finding a set of respective feature values for each sub-window of the divided observation window;

- applying a classification algorithm on all the respective sets of sub-window feature values to induce a probability estimation model, while considering, among others, the labelled outcome;

- storing the probability estimation model; and during a real-time stage: - continuously updating the divided observation window with new real-time patient's monitored data;

- periodically determining a set of real-time feature values for each sub-window of the divided observation window;

- periodically submitting each of said determined real time feature values to the probability estimation model and calculating by said probability estimation model a probability for a hypotensive outcome episode to occur within the target window .

2. A method according to claim 1, further allowing a user to tune a probability threshold, above which an alert is issued.

3. Method according to claim 1, wherein said determination of the probability estimation model involves consideration of inter sub-windows feature values, and wherein said probability calculation by the probability estimation model involves consideration of inter sub-windows feature values.

4. A method according to claim 1, wherein said features are selected from: statistical features, wavelet features, cross correlation features, dynamic features, and distribution features .