Classifications

• • A—HUMAN NECESSITIES
  • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
  • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
  • A61B10/00—Instruments for taking body samples for diagnostic purposes; Other methods or instruments for diagnosis, e.g. for vaccination diagnosis, sex determination or ovulation-period determination; Throat striking implements
• • G—PHYSICS
  • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
  • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
  • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
  • G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
• • G—PHYSICS
  • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
  • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
  • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
  • G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients

Definitions

• the invention relates to the field of clinical decision support where an estimation value of thrombosis risk of a patient is calculated based on patient-specific input features.
• CDSSs Computer-based clinical decision support systems
• Clinical decision support systems have been promoted for their potential to improve the quality of health care by supporting clinical decision making.
• Deep vein thrombosis is a wide spread problem in the western world. Large portions of the population are at increased risk of thrombosis, e.g. the elderly, people who travel, and patients that undergo orthopedic surgery. People at risk can be put on preventive anticoagulant treatment, but the risk of bleeding (1-3% per year), and issues of cost and inconvenience speak against this. It would therefore be desirable to have a more patient- specific measure to estimate the personal thrombosis risk and facilitate an informed choice on whether or not to treat. Unfortunately, with current clinical screening techniques and available methodologies, high risk individuals, which should receive anticoagulants, are not easily recognized and events are not accurately predicted.
• the proposed solution helps the physician to stratify the patients that are treated or examined for conditions that are known to increase thrombosis risk, into high and low risk categories. Specifically, the proposed solution may be used to decide, per patient, whether or not to administer anticoagulant treatment based on estimated thrombosis risk.
• the term "molecular marker" is intended here to include any use of the presence or concentration of a biomolecule or part of a biomolecule, e.g., a protein or a polynucleid acid as an indicator of a patient phenotype. Such presence or concentration may be measured directly in e.g. a blood or tissue sample, or as a (possibly dynamic)
• PCR polymerase chain reaction
• thrombin generation assay a functional test like real-time quantitative polymerase chain reaction (PCR) or the thrombin generation assay.
• At least one molecular marker may be selected from a concentration of coagulation protein FVIII in blood, a concentration of coagulation protein FXI in blood, and a concentration of coagulation protein TFPI in blood. Based on patient datasets obtained from a clinical study, these types of protein concentrations have turned out to serve as reliable indicators of thrombotic risk.
• At least one clinical risk factor may be selected from immobilization within a first predetermined time period, surgery within a second predetermined time period, family history of venous thrombosis, pregnancy or puerperium with a third predetermined time period, current use of estrogens, and obesity.
• the first predetermined time period may correspond to at least three months
• the second predetermined time period may correspond to one month
• the third predetermined time period may correspond to at least three months.
• the estimation value of thrombotic risk may be compared with a
• predetermined threshold value in order to classify the estimation value based on the comparison result.
• a user may be allowed to input or disable the predetermined threshold value.
• the decision support mechanism can be adapted based on the needs of the user (i.e. clinician).
• an optimization mechanism may be provided for applying a learning process through an optimization procedure based on a dataset stored in a database so as to minimize a prediction error. This allows continuous adaptation of the clinical decision support mechanism to new datasets of new patients or to specific datasets of individual patients.
• the dataset may be divided into a training set, a validation set and a test set, wherein the training set and the validation set may be used to select a type of machine learning function and a set of model parameters used for optimizing classifiers, wherein the optimized classifiers may be used for obtaining the patient-specific input features, and wherein the test set may be used for monitoring the estimation value for patients of the test set based on the obtained input features.
• This measure allows specific trimming of the input features of the clinical decision support system to a data set obtained from a specific group of patients to thereby further enhance reliability of risk estimation.
• said processor is adapted to calculate a deep vein thrombosis (DVT) risk score, representing an estimation value of thrombosis risk of a patient, based on clinical risk factors, single nucleotide polymorphisms (SNPs) and protein levels.
• DVT risk score shows significant improvement in terms of sensitivity/specificity over known methods that calculate a DVT risk score without protein levels.
• the apparatus may be implemented as a discrete hardware circuitry with discrete hardware components, as an integrated chip, as an arrangement of chip modules, or as a signal processing device or chip controlled by a software routine or program stored in a memory, written on a computer readable medium, or downloaded from a network, such as the Internet.
• Fig. 1 shows a schematic block diagram of a clinical decision support system according to various embodiments
• Fig. 2 shows a flow diagram of a risk estimation procedure according to a first embodiment
• Fig. 3 shows a flow diagram of a classifier optimization procedure according to a second embodiment
• Fig. 4 shows a schematic representation of a user interface according to a third embodiment
• Figs. 5 A and 5B respectively show a receiver operator curve (ROC) plus 95% confidence interval for thrombosis predicted by a support vector machine with only clinical risk factors as input resulting and a ROC curve plus 95% confidence interval for thrombosis predicted by a classifier with clinical risk factors and protein concentrations as inputs; and
• ROC receiver operator curve
• Figs. 6 A and 6B respectively show a ROC plus 95% confidence interval for thrombosis, predicted within the subgroup of patients with one or more known clinical risk factors present, by a support vector machine with only clinical risk factors as input and a ROC curve plus 95% confidence interval for thrombosis predicted by a classifier with clinical risk factors and protein concentrations as inputs.
• Embodiments are now described based on a computerized clinical decision support system for predicting thrombosis risk based on a combined consideration of clinical risk factors and molecular markers, e.g., protein concentrations.
• Fig. 1 shows a schematic block diagram of a clinical decision support system according to various embodiments, which involves a clinical decision support algorithm and/or software. It comprises data interface (DI) 10 where information about a specific patient is made available to the system, a processor (P) 20 which applies an interpretative algorithm and a user interface (UI) 30 which makes the interpretation of the calculated data available to a user, e.g., a clinician. Furthermore, an optional optimization system may be provided for optimizing classifiers so as to provide a good trade-off between good prediction accuracy and conciseness of the set of input features or parameters for the clinical decision support algorithm.
• the optimization system comprises an optimization unit (O) 40 which may be based on a separate processor running an optimization software or based on a separate software routine controlling the processor 20.
• the optimization unit 40 retrieves data required for optimization from a database (DB) 50.
• DB database
• the data interface 10 may be a classical user interface for allowing interaction between a user and the clinical decision support system, or a direct link to a central computer database or electronic patient record. In either case, the data interface 10 is adapted to collect at least some of the following input features on a patient at the date on which the clinical decision support system is used to assess thrombosis risk:
• immobilization plaster cast, extended bed rest at home for at least 4 days, hospitalization
• last three months e.g. "1" for true, "0" for false
• estrogens oral contraceptives or hormone replacement therapy (e.g. "1" for true, “0” for false));
• body mass index over 30 e.g. "1" for true, “0” for false
• concentration (U/mL) of the coagulation protein FVIII in blood concentration (U/mL) of the coagulation protein FXI in blood; and concentration (ng/ml) of the coagulation protein TFPI in blood.
• the processor 20 calculates a numerical function of the above list of numerical inputs by applying the clinical decision support algorithm.
• This numerical function returns a number, i.e. risk score (R), between zero and one, where zero is the lowest possible thrombosis risk indication and one is the highest.
• This numerical output may be shown directly on the user interface 30 and/or may be compared to a threshold (T) between zero and one. If the risk score exceeds the threshold T, anti-coagulant therapy is indicated for the patient for whom the values have been entered into the calculation. Otherwise, preventive anti-coagulation therapy is indicated as not advisable.
• T which can be set as a fixed value in the system or tuned by the user at the user interface 30, determines the balance between sensitivity and specificity of the clinical decision support system. Low values for T will infer a bias towards the indication of high risk, which leads to few false negatives (high sensitivity) but increases the number of false positives (low specificity or overtreatment). High values for T give the opposite effect and tends to undertreatment.
• the specific choice of T is the responsibility of the user, e.g.
• the clinical decision support system may be implemented as a software application on a computer (system) that can be accessed by a clinician who needs to make a decision about patients' anticoagulation treatment.
• the software application of the clinical decision support system may be integrated (e.g. as a plug-in) in an existing hospital information management system.
• the interpretative clinical decision support algorithm may be a complex mathematical function that takes numerical (or Boolean) values for the above nine input features as input, uses these in a series of non-linear calculations and returns a numerical value between zero and one, where higher values represent a higher risk of thrombosis.
• the numerical function consists of one or a combination of classifier functions that are common in the field of machine learning, such as neural network functions or support vector machines or Bayesian network. These classifiers are optimized by the optimization unit 40 based on the database 50 of subjects, i.e. thrombosis patients and healthy controls for whom numerical values for the aforementioned nine input features are available.
• Optimization of the optimization unit 40 involves tuning the parameters of the classifier functions in such a way that the correlation between calculated risk score on the subjects in the database and recorded occurrence of thrombosis is maximized.
• the optimization process constitutes a significant effort that requires a strong experience in and understanding of the field of machine learning and numerical optimization. The process is further strongly dependent on the quality of the underlying database 50.
• Fig. 2 shows a flow diagram of a thrombosis risk estimation process according to a first embodiment.
• the data interface 10 accesses in step S201 the hospitals electronic patient record (EPR), if present, and reads out the nine patient features that were listed above.
• the user may be requested or allowed to manually enter, e.g. via the user interface 30, numerical values for patient features that are not available from the EPR.
• the data interface 10 checks the entered values for the right numerical format and an error message can be generated if the input format does not match with the required format. In case of a wrong format, the data is converted in step S203 to the numerical formats indicated in the above list, if necessary.
• the user interface 30 may allow the user either to enter a numerical value for the threshold T between zero and one, or to disable the threshold.
• step S204 the procedure checks whether risk calculation has been requested by the user (e.g. through clicking on a respective button at the user interface 30). If not, the systems repeats the above steps S201 to S203 to allow an update of the input features or simply repeats step S204 until risk calculation is requested. I.e., the "No" branch arrow of step S204 can simply point back to the top of step S204 and needs not go back to step S201. If the request is detected in step S204, clinical decision support algorithm is called in step S205 (e.g. by the processor 20) to calculate a risk score based on the input features gathered in the previous steps.
• step S205 clinical decision support algorithm
• step S206 it is checked if the threshold (T) has been enabled. If not, the procedure branches to step S209 and the calculated risk score is shown as a number or another graphical representation e.g. on a computer screen or other output medium of the user interface 30 before the procedure ends in step S210. Otherwise, if the procedure detects in step S206 that the threshold has not been disabled, the risk score is compared in step S207 to the threshold and classified based on the result of comparison. Finally, in step S208 a classification of 'high thrombosis risk' or 'low thrombosis risk' is made visible e.g. on the screen of the user interface 30 dependent on whether the risk score is higher or lower than the threshold. Optionally, a numerical and/or graphical comparison between the threshold value and the risk score should be shown along with the classification.
• the risk score could be calculated continuously (instead of upon request). This could also be done with some of the missing input parameters. In that case, a range of possible risk scores (e.g., indicated by a minimum risk estimation and maximum risk estimation) is provided as output, e.g., based on an uncertainty in the calculation.
• the required data set of the database 50 may be derived from a data collection based on an extensive questionnaire on many potential risk factors for venous thrombosis. More specifically, the data collection may involve information (e.g. clinical risk factors) obtained from a questionnaire and clinical assays (e.g. activity or antigen-based assays of protein concentrations) as described in the respective assay protocols. .
• information e.g. clinical risk factors
• clinical assays e.g. activity or antigen-based assays of protein concentrations
• Machine learning methods are black box methods that exploit the patterns that may be hidden in the numerical values of the data to predict an output. Each method constructs a mathematical function that takes observed quantities (like protein
• Such a function is defined through its structure (e.g. a neural network function) and the numerical value of the function parameters (e.g. the weights in a neural network).
• the combination of function structure, parameter values and numerical inputs produce an output feature which may be binary (e.g. thrombosis vs. no thrombosis), or continuous (e.g. probability of thrombosis).
• the specific type of method that is used in the second embodiment is the support vector machine (SVM), an often used method in the field of machine learning (see e.g.
• Fig. 3 shows a flow diagram of an optimization process according to a second embodiment.
• a classifier is a specific class of black box model, the output of which is the class or label of a data element, where each element is described by a number of numerical features.
• the data elements in the present embodiments are human subjects for whom a number of clinical features are known through measurement or anamnesis.
• the class is binary: thrombosis patient or control subject.
• the classifier is trained on the dataset of the database 50 which contains each participant's numerical features and the corresponding label.
• step S301 the dataset of the database 50 is divided in step S301 into three equally sized sets, called training set, validation set and test set, each containing the same ratio of cases to controls.
• the training set is used for training or parameter tuning, i.e. search for that set of parameter values that minimizes the prediction, or in this case classification error.
• Most machine learning methods suffer from so-called 'overfitting', where the method's performance on the training set is much better than its performance on new data that has not been used for training. Therefore, in step S303, a separate validation set is used to test whether such over- fitting occurs.
• step S304 The combination of training and validation data allows to find that type of machine learning function and choice of model parameters that is able to grasp the true pattern that hides in the (training) data, yet is still sufficiently general to predict well on the separate validation data and thus on future data as well.
• the thus optimized classifiers are used in step S304 to make a prediction on each of the patients in the test set, which has remained unused throughout the foregoing optimization steps.
• the quality of this prediction (e.g. in terms of sensitivity and specificity) is the final test of the validity of the selected classifier.
• the test set is selected at random to obtain solid statistics.
• the steps S301 to S303 described the selection of an optimal classifier based on a train and validation subset of a database.
• Through permutation of the subjects in the train and the validation set (swapping patients between the two sets) in step S305 it is possible to create an ensemble of classifiers, each classifier corresponding to one specific permutation of train and validation subjects.
• Such an ensemble is used as a voting system. This means that each classifier in the ensemble assigns a label to the same object, e.g.
• the label that turns up most often is assumed to be the correct one, and the fraction of votes that support this label are used as a confidence score: if all classifiers in the ensemble vote for thrombosis, it is 100% sure that the participant will get thrombosis, whereas a fifty-fifty distribution of the votes makes the classification no better than a coin flip.
• the risk score (R) is compared to a threshold (T), where a score that exceeds the threshold indicates a case and a score below the threshold indicates a control subject.
• step S305 When the optimal classifier on the complete set of features has been found in step S305, the relative importance of each input feature in the classifier is analyzed in step S306.
• the selected subjects in the train and validation set are now used to select those features that contribute most to a correct classification.
• the following input reduction procedure is executed in step S306 for each of the optimized classifiers:
• the above reduction procedure is used to deduce a selection of overall most predictive features. It is performed for each aforementioned (random) division of the complete database into a train, validation and test set. In step S307, for each division, the classifier is reduced to ten input features, and each remaining input feature is marked. Then, in step S309, the number of times each input feature remains in the 'top ten' is counted and this count is used to rank the input features from most predictive (part of the top ten most often) to least predictive. Finally, the most predictive input features are used for risk calculation in the clinical decision support algorithm of the processor 20 and the procedure ends in step S310.
• the optimization procedure of the second embodiment can be used to regularly update the clinical decision support algorithm of the processor 20 based on new patient data in the database 50.
• Fig. 4 shows a schematic representation of a front view of the user interface 30 of Fig. 1.
• the patient name (PN) and its identification number (ID) is indicated as "Jane Doe” and "099812".
• PN patient name
• ID identification number
• nine input features are designated and their actual binary values ("0" or "1") of the above patient are indicated on the right side beneath the designation.
• the first six input features are the clinical risk factors indicating recent surgery (RS), obesity (O), family history (FH), Immobility (I),
• the last three input features are the concentration levels of coagulation proteins Factor VIII (FVIII), Factor XI (FXI) and tissue factor pathway inhibitor (TFPI).
• T the currently set threshold level
• DA tissue factor pathway inhibitor
• T the currently set threshold level
• a button (CAL) for activating or triggering a risk calculation by the processor 20 is shown.
• RS calculated risk score
• RV graphical visualization
• STR stratification
• the bar which indicates the current risk score on the risk scale is qualified as low risk (LR).
• LeidenThrombophilia Study (as described for example in van der Meer et al.: “The LeidenThrombophilia Study (LETS)", Thromb Haemost. 1997;78(1):631-5) in order to find a combination of known biomarkers that is able to distinguish thrombosis patients from healthy controls.
• Focus was directed at two different types of patient features, i.e. coagulation protein concentrations in blood and clinical risk factors that are known to relate to thrombosis. It could be shown that the predictive power of clinical risk factors alone, either as a simple risk factor count or used in a machine learning approach, can be improved by incorporation of measured coagulation protein concentrations.
• Figs. 5A and 5B show respective diagrams with a receiver operator curve (ROC) plus 95% confidence interval for thrombosis predicted by a support vector machine with only clinical risk factors as input resulting in an area under the ROC curve (AUC) of 0.72 (0.68 - 0.77) (Fig. 5A) and a ROC curve plus 95% confidence interval for thrombosis predicted by a classifier with clinical risk factors and protein concentrations as inputs resulting in an AUC of 0.78 (0.74 - 0.83) (Fig. 5B).
• the ROC curves plot the true positive rate (vertical axis) against the false positive rate (horizontal axis) for different threshold values.
• AUC area under the ROC curve
• a second example relates to input feature reduction.
• the determined most influential protein in thrombosis classification was coagulation factor VIII, followed by factor XI and TFPI (cf. Table 1 below).
• Classification with all clinical risk factors (for which no measurement is necessary) and these three protein concentrations achieves almost equivalent classification at AUC of 0.77.
• the improvement is especially clear in the increased risk population, here defined as those subjects showing one or more known clinical risk factors.
• Figs. 6A and 6B show the ROC plus 95%> confidence interval for thrombosis, predicted within the subgroup of patients with one or more known clinical risk factors present, by a support vector machine with only clinical risk factors as input resulting in an AUC of 0.67 (0.60 - 0.75) (Fig. 6A), and a ROC curve plus 95% confidence interval for thrombosis predicted by a classifier with clinical risk factors and protein concentrations as inputs resulting in an AUC of 0.75 (0.69 - 0.81) (Fig. 6B).
• the use of the three protein concentration values allows a further stratification of this risk group with an ROC score of 0.75 versus 0.67 based on the use of clinical risk factors alone (number of co-occurring factors or knowledge of which factor is present).
• Table 1 shows a list of classifier features, sorted by the percentage of classifiers (based on different random choices of validation set) that retain the feature in the 10 features that are pruned last.
• the risk of deep vein thrombosis has been evaluated by using information from the MEGA (Multiple Environment and Genetic Assessment of risk factors for venous thrombosis) study and the Leiden Thrombophilia Study (LETS). Both are case-control studies that were set up to identify risk factors for venous thrombosis that have been performed in the Netherlands (Blom, 2005, van der Meer FJ, Koster T, Vandenbroucke JP, Briet E, 1997).
• a plethora of variables, ranging from coagulation protein levels to environmental thrombotic risk factors and genetic thrombophilia has been taken from patients with venous thrombosis and controls.
• a neural networks approach see e.g.
• Kuncheva, 2004 has been used in the MEGA study to estimate potential risk factors for Deep Vein Thrombosis (DVT) and their predictive value in one integrated approach.
• the identified combinatory risk score is validated in an internal cross- validation on the MEGA study and in an independent validation on the LETS study.
• immobilization because of plaster cast, leg injury in the past 3 months, cancer in the period from five years before to six month after the index date and travel for more than four hours in the past 2 months.
• the other considered risk factors were part of the initial study as well: immobilization because of extended bed rest at home for at least 4 days, hospitalization), surgery, a family history of venous thrombosis (considered positive if at least 1 parent, brother, or sister experienced venous thrombosis, pregnancy or puerperium within 3 months before the index date, or use of estrogens (oral contraceptives or hormone replacement therapy) at the index date and the presence of obesity, determined as a body mass index of 30 kg/m2 or higher).
• the considered protein levels are a subset of the proteins that were included before (because of a more limited set of measurements performed in the MEGA study). They are: anti-thrombin (AT), prothrombin (factor II), factor 7 (FVII), FVIII, FIX, FX, FXI, fibrinogen and protein C (all activity measurements) and protein S (antigen measurement).
• the LETS study includes four less clinical risk factors than the MEGA study, as described above with respect to the clinical risk factors.
• the cross-validation as performed in the previous paragraph has been repeated without these four risk factors and under the exclusion of cancer patients, who had been excluded from the LETS study as well.
• the AUCs on the reduced MEGA study are 0.84, 0.80 and 0.74, in the same order as in the last paragraph.
• one risk score on the reduced MEGA study (without divisions into train and test set as would be necessary in a cross-validation) was derived and applied this risk score without adaptation to the individuals of the LETS study.
• the resulting AUCs were 0.82, 0.79 and 0.74, showing that the proposed risk score can be applied on an independent study with little loss of performance, and the improvement due to the proposed inclusion of protein levels holds in an external validation.
• an apparatus and method have been described for clinical decision support to identify patients at high risk of thrombosis based on a combination of clinical risk factors and molecular markers, e.g., protein concentrations. These clinical risk factors and molecular markers are combined in a machine learning based algorithm which returns an output value, relating to an estimated risk of a thrombosis event in the future.
• clinical risk factors and molecular markers e.g., protein concentrations.
• concentration of protein Z, C4B binding protein, fibrinogen, TAFI, Factor II, V, VII, IX, X, XII or XIII, antithrombin, protein C, protein C inhibitor, protein S or other markers may be selected as decisive input features.

Landscapes

• Engineering & Computer Science (AREA)
• Health & Medical Sciences (AREA)
• Medical Informatics (AREA)
• Public Health (AREA)
• Data Mining & Analysis (AREA)
• Pathology (AREA)
• Biomedical Technology (AREA)
• General Health & Medical Sciences (AREA)
• Primary Health Care (AREA)
• Epidemiology (AREA)
• Databases & Information Systems (AREA)
• Life Sciences & Earth Sciences (AREA)
• Surgery (AREA)
• Veterinary Medicine (AREA)
• Animal Behavior & Ethology (AREA)
• Molecular Biology (AREA)
• Heart & Thoracic Surgery (AREA)
• Investigating Or Analysing Biological Materials (AREA)

Priority Applications (6)

Applications Claiming Priority (2)

Publications (1)

Family

ID=49956255

Family Applications (1)

Country Status (7)

Cited By (2)

Families Citing this family (17)

Citations (2)

Family Cites Families (5)

• 2013
  • 2013-10-17 CN CN201380055556.4A patent/CN104756117B/zh active Active
  • 2013-10-17 EP EP13821154.5A patent/EP2912584B1/en active Active
  • 2013-10-17 BR BR112015009056A patent/BR112015009056A2/pt not_active IP Right Cessation
  • 2013-10-17 RU RU2015119520A patent/RU2682622C2/ru not_active IP Right Cessation
  • 2013-10-17 US US14/434,286 patent/US20150278470A1/en not_active Abandoned
  • 2013-10-17 WO PCT/IB2013/059424 patent/WO2014064585A1/en not_active Ceased
  • 2013-10-17 JP JP2015538597A patent/JP6335910B2/ja not_active Expired - Fee Related

Patent Citations (2)

Non-Patent Citations (7)

Also Published As

Similar Documents

Legal Events