SYSTEM AND METHOD FOR SEPSIS PREDICTION
BACKGROUND
[0001] The subject matter disclosed herein generally relates to predictions and prognostics of diseases in a subject. More specifically, the subject matter relates to systems and methods for predicting sepsis conditions such as of a hospitalized patient.
[0002] Sepsis is more commonly called a blood stream infection or blood poisoning. It is generally caused by presence of bacteria, infectious organisms, or their toxins in the blood or other tissues of the body. Sepsis often occurs in patients suffering from systemic inflammatory response syndrome- (SIRS), as a result of a number of conditions such as surgery, trauma, burns, pancreatitis and other non-infectious events. Sepsis may be associated with clinical symptoms of systemic illness, such as fever, chills, malaise, low blood pressure, and mental status changes. Sepsis is a serious situation associated with high morbidity and mortality rates. Treatment depends on the type of infection, but usually begins with antibiotics or similar medications.
[0003] It is currently estimated that sepsis is the tenth leading cause of death in the United States with an overall rate of mortality as high as 35%, an estimated 750,000 cases per year, and the annual cost to treat sepsis in the order of billions of dollars.
[0004] Early diagnosis of sepsis is difficult. Diagnosing sepsis sufficiently early allows for more effective intervention and prevention. Most existing sepsis scoring systems or predictive models predict only the risk of late-stage complications, in patients with severe sepsis or septic shock. Such techniques, however, do not predict the development of sepsis itself, especially in early stages. There is a need for identifying those with SIRS who would develop sepsis.
BRIEF DESCRIPTION
[0005] In accordance with one aspect of the present technique, a computer implemented method is disclosed. The method includes providing a model configuration and obtaining clinical data about a patient. The model configuration includes definition of a feature vector,

a sepsis model, and a confidence function and the clinical data includes stored clinical data and measured clinical data. The method further includes estimating a plurality of feature parameters of the feature vector based on the measured clinical data and determining proxy data based on the stored clinical data. The proxy data corresponds to at least one missing parameter of the feature vector, if the at least one missing parameter is different from the plurality of feature parameters. The method also includes predicting a sepsis probability value based on the sepsis model, the feature vector, and the proxy data and generating a confidence value based on the sepsis probability value and the confidence function. The method further includes identifying the measured clinical data required to determine the at least one missing parameter, if the confidence value is less than a confidence threshold value and generating a sepsis condition prediction based on the sepsis probability value, with the confidence value indicative of reliability of the sepsis condition prediction.
[0006] In accordance with another aspect of the present technique, a system is disclosed. The system includes at least one processor and a memory coupled to communications bus and a signal acquisition module configured to provide a model configuration and obtain clinical data of a patient. The model configuration includes definition of a feature vector, a sepsis model, and a confidence function and the clinical data includes stored clinical data and measured clinical data. The signal acquisition module is further configured to estimate a plurality of feature parameters of the feature vector based on the measured clinical data and determine proxy data corresponding to at least one missing parameter of the feature vector, based on the stored clinical data, if the at least one missing parameter is different from the plurality of feature parameters. The system also includes a sepsis prediction module communicatively coupled to the signal acquisition module and configured to generate a sepsis probability value based on the sepsis model, the feature vector, and the proxy data. The sepsis prediction module is further configured to generate a sepsis condition prediction based on the sepsis probability value. The system also includes a confidence generator module communicatively coupled to the sepsis prediction module and the signal acquisition module configured to generate a confidence value based on the sepsis probability value and identify the measured clinical data required to determine the at least one missing parameter, if the confidence value is less than a confidence threshold value. The confidence value

generated by the confidence generator module is indicative of reliability of the sepsis condition prediction. At least one of the signal acquisition module, the sepsis prediction module, and the confidence generator module is stored in the memory and executed by the at least one processor.
[0007] In accordance with another aspect of the present technique, a non-transitory computer readable medium encoded with a program to instruct at least one processor is disclosed. The program instructs the at least one processor to provide a model configuration and obtain clinical data of a patient. The model configuration includes definition of a feature vector, a sepsis model, and a confidence function and the clinical data includes stored clinical data and measured clinical data. The program further instructs the at least one processor to determine a plurality of feature parameters of the feature vector based on the measured clinical data. The program also instructs the at least one processor to estimate proxy data corresponding to at least one missing parameter of the feature vector, based on the stored clinical data, if the at least one missing parameter is different from the plurality of feature parameters. The program also instructs the at least one processor to predict a sepsis probability value based on the sepsis model, the feature vector, and the proxy data and generate a confidence value based on the sepsis probability value and the confidence function. The program further instructs the at least one processor to identify the measured clinical data required to determine the at least one missing parameter, based on the confidence value, if the confidence value is less than a confidence threshold value. The program instructs the at least one processor to generate a sepsis condition prediction based on the sepsis probability value, with the confidence value indicative of reliability of the sepsis condition prediction.
DRAWINGS
[0008] These and other features and aspects of embodiments of the present invention will become better understood when the following detailed description is read with reference to the accompanying drawings in which like characters represent like parts throughout the drawings, wherein:

[0009] FIG. 1 is a diagrammatic illustration of a system for predicting sepsis conditions of a patient in accordance with an exemplary embodiment;
[0010] FIG. 2 illustrates a schematic flow diagram illustrating processing of clinical data in accordance with an exemplary embodiment;
[0011] FIG. 3 illustrates a confidence, function in accordance with an exemplary embodiment;
[0012] FIG. 4 illustrates a tree model in accordance with an exemplary embodiment; and
[0013] FIG. 5 is a flow chart illustrating a method for sepsis prediction of a patient in accordance with an exemplary embodiment.
DETAILED DESCRIPTION
[0014] Embodiments of a system and a method for predicting a sepsis condition during treatment of a patient are disclosed. Treatment refers to some form of medical diagnostics such as by a medical professional and in some examples the treatment is at a medical facility such as a hospital, lab, medical clinic, physician's office and the like. Specifically, in certain. embodiments, clinical data related to a patient is obtained. The patient or subject is the human or animal that is undergoing the treatment. The clinical data typically includes stored clinical data and measured clinical data. A plurality of feature parameters of a feature vector are determined based on the measured clinical data. In one example when at least one of the feature parameter of the feature vector is missing, the stored clinical data is used to determine proxy data corresponding to the at least one missing feature parameter of the feature vector. A sepsis probability value is determined based on aspects such as the feature vector, a sepsis model and the proxy data. A sepsis condition prediction is determined based on the sepsis probability value. A confidence value is determined based on the sepsis probability value and a confidence function. The measured clinical data required to determine the at least one missing feature parameter is identified based on the confidence value.
[0015] FIG. 1 is a diagrammatic illustration of a system 100 for predicting sepsis condition of a patient 102 in accordance with an exemplary embodiment. The system 100

includes a.signal acquisition module 106, a sepsis prediction module 108, and a confidence generator module 110. The system 100 also includes at least one processor 124 and memory 120. The system 100 includes a communication bus 122 providing communication between the signal acquisition module 106, the sepsis prediction module 108, the confidence generator module 110, the memory 120 and the at least one processor 124.
[0016] In operation, the signal acquisition module 106 receives clinical data 104 from the patient 102. The clinical data 104 typically includes a plurality of chart events corresponding to the patient 102, readings from a plurality of lab tests conducted on the patient 102, a plurality of waveforms acquired from the patient and various patient demographic parameters. The signal acquisition module 106 receives the clinical data as measured clinical data and stored clinical data. The measured clinical data corresponds to clinical data of the patient 102 acquired during treatment such as hospitalization, and in one example represents contemporaneous clinical data. The stored clinical data corresponds to clinical data corresponding to patients acquired in the past, and/or represents historical clinical data corresponding to the patient 102 in previous treatments or health evaluations. The measured clinical data and the stored clinical data may be stored in memory storage and accessible for use in the treatment. In one embodiment, the signal acquisition module. 106 is a computer program stored in memory 120 and executable by the at least one processor 124. In another embodiment, the signal acquisition module 106 is standalone hardware module configured to acquire the clinical data 104.
[0017] The signal acquisition module 106 also receives a model configuration 132 required to perform sepsis prediction for the patient 102. The model configuration includes definition of a feature vector 126, a sepsis model 128, and a confidence function 130. The definition of the feature vector 126 includes items such as a list of feature parameters and an arrangement of the feature parameters in the feature vector. The definition of the feature vector 126 also includes computer implementable mathematical formulae required to determine the feature parameters of the feature vector. The term feature parameter used herein refers to a parameter that can be computed based on the clinical data that are useful in predicting sepsis condition in a patient. The feature parameters of the feature vector are determined based on the clinical data 104. The confidence function is any suitable

mathematical function used for computing a reliability value corresponding to the sepsis condition prediction. The model configuration further specifies a method for determining the proxy data. The term proxy data refers to an estimate of at least one missing parameter or an estimate of an alternate feature parameter that may be substituted for at least one missing parameter. In one embodiment, the model configuration 132 is provided by an operator of the system. In another embodiment, the model configuration 132 is stored in the memory and is retrieved by the signal acquisition module. In one example the model configuration 132 is a computer program executed by the signal acquisition module 106, instantiating the feature vector 112.
[0018] The sepsis model refers to one of a physical or mathematical model used to predict a sepsis condition in a patient. The sepsis model includes one or more model parameters and a model threshold value. The confidence function includes one or more function parameters. The model parameters are used by the sepsis prediction module 108 and the function parameters are used by the confidence generator module 110. In one embodiment, the model configuration is stored in memory 120 and retrieved by the signal acquisition module 106. In certain embodiments, the sepsis model parameters and the function parameters are included in the model configuration and are received by the signal acquisition module 106.
[0019] In one exemplary embodiment, the model configuration includes a structure of the sepsis model. In another embodiment, the sepsis model may be selected by the signal acquisition module 106. In this embodiment, one or more model parameters of the sepsis model and the model threshold value are estimated by the signal acquisition module 106 based on the stored clinical data.
[0020] According to one example, the sepsis model is evaluated based on a cross validation technique. The cross validation technique also provides a receiver operating characteristic (ROC) as a graph of true positive rate (also referred as TPR, or sensitivity, or recall) as a function of false positive rate (also referred as FPR, or fall out). The false positive rate is equal to (1 - specificity). The true positive rate represents a number of true positives among the total positive decisions and the true negative rate represents a number of

true negatives among the total negatives decisions. The model threshold value is determined based on a true positive rate value and a false positive rate value. The true positive rate value and the false positive rate value used for determining the model threshold value may be specified by a user. Alternatively, desired values of the true positive rate and the false positive rate are chosen or selected by the operator. In another embodiment, acceptable values of the true positive rate value and the false negative value may be used in determining the model threshold value. If the model threshold value is not available for a given combination of true positive rate and false positive rate, modification in the model configuration is needed. At least one of the sepsis model and the feature vector are modified and the model parameters are re-estimated. The process is repeated until the model threshold value is obtained for the specified values of the true positive rate and the false positive rate. It should be noted herein that, in some embodiments, additional performance criteria which may be subjective in nature are also considered while determining the model threshold .value.
[0021] In another exemplary embodiment, the model configuration includes a structure of the confidence function in the form of a mathematical equation in a computer program of the model configuration. The confidence function includes the function parameters and a distance parameter. The distance parameter is a difference value between a class probability value generated from the sepsis model and the model threshold value. In this embodiment, the signal acquisition model 106 also determines the function parameters of the confidence function based on the stored clinical data. A plurality of values of the distance parameters and corresponding class category information are used to determine the function parameters of the confidence function. The term class category, used herein, refers to a 'sepsis class' category or a 'not sepsis class' category. It should be noted that a sepsis model is used in this embodiment to generate the class probability value. The sepsis model provided in the model configuration or the estimated sepsis model based on the model configuration may be employed.
[0022] The sepsis prediction module 108 is communicatively coupled to the signal acquisition module 106 and configured to receive the feature vector 112. The sepsis prediction module 108 includes the sepsis model and determines a sepsis probability value 114 based on the feature vector 112 received from the signal acquisition module 106. The

sepsis probability value 114 is indicative of a probability of the patient 102 developing sepsis condition during treatment. The sepsis prediction module 108 predicts a sepsis condition based on the sepsis probability value and the model threshold value. The sepsis condition prediction is a binary variable having values 'zero' and 'one' representing 'not sepsis class' category and 'sepsis class' category respectively. In one embodiment, the sepsis prediction module 108 is stored in the memory 120 and executable by the at least one processor 124. In another embodiment, the sepsis prediction module 108 is a standalone hardware module configured to determine the sepsis probability value 114.
[0023] The confidence generator module 110 is communicatively coupled to the signal acquisition module 106 and the sepsis prediction module 108 by the communications bus 122. The confidence generator module 110 generates a confidence value 116 based on the sepsis probability value 114 and the confidence function. The confidence value 116 generated by the confidence generator module 110 provides a confidence measure on the sepsis condition prediction. In an exemplary embodiment, the confidence value 116 is used for improving the quality of the sepsis probability value 114. The confidence generator module 110 identifies measured clinical data required to determine at least one missing parameter of the feature vector 112. The identified measured clinical data may be obtained by one or more lab tests and at least one missing parameter may be determined based on the identified measured clinical data. In one embodiment, the confidence generator module 110 is stored in the memory 120 and executable by the at least one processor 124. In another embodiment, the confidence generator module 110 is a customized hardware module configured to generate the confidence value 116.
[0024] In an embodiment of the presented technique, the confidence generator module 110 is further configured to generate a visual information 118 based on the plurality of feature parameters of the feature vector 112, the sepsis probability value 114, and the confidence value 116 in a suitable format. The visual information 118 represents relations among the plurality of feature parameters of the feature vector 112 with reference to a sepsis prediction value and the confidence value 116. The visual information 118 is presented to an operator through a display device such as a computer screen in the format of a graph, or a tree. The visual information 118 provides transparency into the intermediate steps in the

prediction of the sepsis probability value 114. In one exemplary embodiment based on a classification tree, the display is in the form of a tree with nodes representing feature parameters of the feature vector 112 and the links representing intermediate steps in the prediction of the sepsis probability value 114 based on the feature vector 112. In another exemplary embodiment based on a neural network, the display may be in the form of a path from input node to an output node through one or more hidden nodes. The weights associated with the neural network are used to highlight a path providing transparency into the intermediate steps in the prediction of the sepsis probability value 114 based on the feature vector 112.
[0025] The at least one processor 124 includes at least one arithmetic logic unit, a microprocessor, a general purpose controller or a processor array to perform the desired computations or run the computer program. In one embodiment, the functionality of the at least one processor 124 may be limited to generate the feature vector 112. In another embodiment, the functionality of the at least one processor 124 may be limited to generating the sepsis probability value 114. In another embodiment, the functionality of the at least one processor 124 is limited to generating the confidence value 116. In some exemplary embodiments, functionality of the at least one processor would include one or more of the functions of the signal acquisition module 106, the sepsis prediction module 108, and the confidence generator module 110. While the processor 124 is shown as a separate unit, there can be a processor co-located or integrated in one or more of the modules 106, 108, 110. Alternatively, the processor 124 can be local or remote, such as a central server or cloud based, with the communications bus 122 can be wired, wireless or a combination thereof.
[0026] The memory 120 may be a non-transitory storage medium. For example, the memory 120 may be a dynamic random access memory (DRAM) device, a static random access memory (SRAM) device, flash memory or other memory devices. In one embodiment, the memory 120 may include a non-volatile memory or similar permanent storage device, media such as a hard disk drive, a floppy disk drive, a compact disc read only memory (CD-ROM) device, a digital versatile disc read only memory (DVD-ROM) device, a digital versatile disc random access memory (DVD-RAM) device, a digital versatile disc rewritable (DVD-RW) device, a flash memory device, or other non-volatile storage devices.

In one specific embodiment, a non-transitory computer readable medium may be encoded with a program to instruct the at least one processor 124 to determine the sepsis probability value.
[0027] FIG. 2 illustrates a schematic flow diagram 200 illustrating processing of the clinical data 104 in accordance with an exemplary embodiment via the signal acquisition module 106, sepsis prediction module 108 and the confidence generator module 110 from FIG. 1. The clinical data 104 includes measured clinical data 202 and stored clinical data 204. The measured clinical data 202 and the stored clinical data 204 includes, for example, a plurality of chart events 206, readings of a plurality of lab tests 208, a plurality of parameters of patient demographic data 210, and a plurality of 1CU specialty attributes 214. The plurality of chart events 206 include but not limited to heart rate (HR) readings, respiration rate (RR) readings, peripheral capillary oxygen saturation (Sp02) readings, systolic blood pressure value (SBP), central venous oxygen saturation level (Scv02) readings and temperature readings. The readings of the plurality of lab tests 208 include but are not limited to white blood cell count (WBC) test, and lactate test. The parameters of the patient demographic data 210 include but are hot limited to family history, age, gender, and weight. In other embodiments, the measured clinical data 202 and the stored clinical data 204 may include hospitalization treatments and medical history 212. The measured clinical data 202 and the stored clinical data 204 typically include readings from a plurality of invasive tests and non-invasive tests. In some embodiments, the measured clinical data 202, and the stored clinical data 204 are taken from a plurality of waveforms acquired from the patient.
[0028] The processing of the clinical data 104 one embodiment is based on the model configuration 132 specified with reference to the system 100 of FIG. 1. The model configuration 132 includes the feature vector 112, a sepsis model 128 and a confidence function 130. The feature vector 112 in this example includes a plurality of feature parameters 222 and at least one missing parameter 224. In some embodiments, the model configuration 132 also includes an imputation method 218. In an alternate embodiment where the imputation method is not used, the model configuration 132 includes a surrogate substitution method 230.

[0029] As shown in FIG. 1, the signal acquisition module 106 determines the feature vector 112 based on the clinical data. Specifically, in an exemplary embodiment, the signal acquisition module 106 determines the plurality of feature parameters 222 of the feature vector 112 based on the measured clinical data 202.
[0030] In some embodiments, one or more feature parameters of the feature vector 112 may not be computable based on the measured clinical data 202. In those embodiments, the signalacquisition module 106 also determines proxy data 236 corresponding to the at least one missing parameter 224 of the feature vector 112 based on the stored clinical data 204. The term proxy data refers herein to an estimate of the at least one missing parameter 224 or to an estimate of an alternate feature parameter that may be substituted for the at least one missing parameter 224. The estimation or interpolation provides a mechanism to provide calculated proxy data 236 as a substitute from measured data.
[0031] The plurality of feature parameters 222 derived from the measured clinical data 202 include one or more of a heart rate, a respiration rate, a Sp02 value, Scv02 value, a temperature value, an age, a weight value, and a white blood cell count. In embodiments of the present technique,, some of the plurality of feature parameters 222 may be derived from non-invasive tests and a few of the plurality of feature parameters 222 may be determined based on the invasive laboratory tests. In one embodiment, the plurality of feature parameters 222 are a plurality of statistical parameters based on at least one of a plurality of heart rate values, a plurality of respiration rate values, a plurality of blood pressure values, a plurality of pulse oximeter oxygen saturation values, and a plurality of body temperature values. The plurality of feature parameters 222 also include standard scores such as SOFA (Systemic Organ Failure Assessment), SAPS (Simplified Acute Physiology Score), and SIRS (Systemic Inflammatory Response Syndrome).
[0032] In another embodiment, the plurality of feature parameters 222 may be an average, a trend, and a standard deviation value of a plurality of numerical values of any of the parameter of the measured clinical data 202 of the patient. For example, one of the plurality of feature parameters 222 of the feature vector 112 may be determined as an average of a plurality of heart rate values. As another example, one of the plurality of feature

parameters 222 of the feature vector 112 may be determined as a standard deviation of a plurality of respiration rate values.
[0033] The proxy data 236 is determined based on the stored clinical data 204, wherein the stored data may be historical data that does not include measured clinical data of the patient or inadequate data of the patient to derive a statistically acceptable value. In an exemplary embodiment, the proxy data is determined based on the imputation method 218 as explained herein. The imputation method determines the proxy data 236 as an estimate of the at least one missing parameter 224 of the feature vector 112. The estimate of the at least one missing parameter is a statistical parameter of the stored clinical data 204. In one embodiment, the imputation method determines a global mean of a plurality of numerical values corresponding to a feature parameter of the feature vector 112. In another embodiment, the imputation method determines a global median of a plurality .of numerical values corresponding to a feature parameter of the feature vector 112. In another embodiment, the imputation method is based on a predefined groups of patients based on demographic parameters. For example, the patients may be grouped into a plurality of non-overlapping categories based on age. A first category may include patient information in the age group of twenty one years to thirty years. A second category may include patient information in the age group of thirty one years to forty years. If the age of a patient age is twenty six years, an estimate of the at least one missing parameter is determined based on the first category. If blood pressure value of the patient is to be determined as an estimate of the at least one missing parameter, a statistical average of a plurality of blood pressure values corresponding to a plurality of patients belonging to the first category is determined.
[0034] In another exemplary embodiment, the proxy data 236 corresponding to the at least one missing parameter 224 is determined based on the surrogate substitution method 230. The surrogate substitution method 230 is used as an alternative to the imputation method 218. The term surrogate feature used herein, refers to an alternate feature parameter that may be used when a feature parameter of the feature vector 112 is missing. As an example, the feature vector 112 may have respiration rate and heart rate as two feature parameters with the respiration rate as a surrogate feature for the heart rate. If heart rate is the missing parameter 224, an estimate of the heart rate value is not used as the proxy data

236 in the surrogate substitution method. Instead, respiration rate value is used as the proxy data during the computation of the sepsis probability value. The respiration rate value may be determined based on the measured clinical data 202.
[0035] The technique of determining surrogate features is explained herein with respect to a tree based model. In a binary decision tree, each node of the tree based model is associated with a splitting rule for selecting one of the two child nodes. The splitting rule is based on a feature parameter of the feature vector 112 and an associated feature threshold value. The feature parameter among the plurality of feature parameters 222 that determines a best splitting rule is associated with each of the node of the tree based model. In the context of surrogate substitution method, the feature parameter determining the best splitting rule is referred as a main feature. Further, an additional feature parameter among the plurality of feature parameters 222 that determines a second best splitting rule at the node is determined. The additional feature parameter is referred herein as a surrogate feature. It should be noted herein that each node may have a plurality of splitting rules and corresponding surrogate features. The assignment of the plurality of surrogate features has an inherent ordering based on a quality of the splitting rule. The quality of splitting rule may be qualitatively determined by suitable techniques such as cross entropy, Gini coefficient, and information gain. In general, one or more surrogate features may be determined at each node of the tree based model, the determination of the surrogate features may be performed by the signal acquisition module of FIG. 1. The surrogate feature is determined based on a similarity criteria and the surrogate feature is similar to the missing parameter of the, feature vector 112.
[0036] The exemplary embodiment of determining proxy data corresponding to the at least one missing parameter using the surrogate substitution method is further explained. In a tree based model, when a main feature of a node is missing, the surrogate substitution method 230 selects the first surrogate feature as a substitute for the at least one missing parameter 224 at the node. The surrogate substitution method 230 selects a second surrogate feature as a substitute for the at least one missing parameter 224 when the main feature and the first surrogate features are missing at the node. For example, for a node in the tree model, a parameter denoted by letter A may be the main feature, and two alternate parameters denoted by letters B and C may be surrogate features in that order. In the absence

of the parameter A, the surrogate feature B is used for branching decision at the node. Further, in the absence of parameters A and the surrogate feature B, the surrogate feature C is used for branching decision at the node. The surrogate features B and C are referred herein as the proxy data 236 corresponding to the feature parameter A.
[0037] The sepsis model 128 is used to process the feature vector 112 and the proxy data 236 to generate the sepsis probability value 114. In one embodiment, the sepsis model 128 is a tree based model such as classification tree and a conditional inference tree. In another embodiment, the sepsis model 128 is a neural network model. Other sepsis models such as logistic equation model and linear model may be used in other exemplary embodiments. The list of model is not exhaustive and does not exclude other similar types of model suitable for the estimation of sepsis condition in a patient. The sepsis model 128 includes one or more model parameters 232. The sepsis probability value 114 is determined using the model parameters 232. For example, if the sepsis model 128 is a linear model, the model parameters 232 are weight coefficients and the sepsis probability value 114 is a weighted sum of the feature parameters of the feature vector 112 weighted by of the model parameters 232. The model threshold value is used to determine class categories with reference to the range of the sepsis probability value 114. In one example, the sepsis probability value 114 has a range of [0 1] and the model threshold value is 0.55. In one example, the sepsis probability values greater than or equal to zero and less than 0.55 determine the 'not sepsis class' category. The sepsis probability values equal to or greater than 0.55 and smaller than or equal to 1 determine the 'sepsis class' category.
[0038] ■ The confidence value 116, generated using the confidence function 130 based on the sepsis probability value 114, is further explained herein. The confidence function 130 may be the computer implemented mathematical function suitable for generating the confidence value 116 corresponding to the sepsis probability value 114 generated by the sepsis model 128. The confidence value 116 indicative of a distance of the sepsis probability value 114 from a boundary between the two class categories representing the 'sepsis class* and the 'not sepsis class'. In one embodiment, the confidence function 130 is a sigmoid function. In another embodiment, the confidence function 130 is based on a logistic regression technique. In another embodiment, the confidence function 130 includes a

significance test on a plurality of leaf nodes in a tree model. The confidence function 130 includes one or more function parameters 234. In embodiments where the model configuration 132 specifies the structure of the confidence function, the one or more function parameters 234 are estimated based on the stored clinical data 204. A least squares method or a regression method may be used to estimate the function parameters 234.
[0039] The sigmoid function f(d,fi) is given by:
where, d is a distance parameter, fi is a scalar value representing the function parameter of the sigmoid function. The distance parameter d, indicative of a difference of the predicted class probability from the decision boundary, is given by:
where, p the sepsis probability value 114 and T is the model threshold value. The confidence value 116 corresponding to the sepsis prediction is defined as:
i.e., the confidence value 116 is determined asf(d,fi) when the sepsis condition is predicted and the confidence value 116 is determined as equal to (1-ffd.fi)) when the sepsis condition is not predicted.
[0040] FIG. 3 illustrates a graph 300 of a confidence function in accordance with an exemplary embodiment. In the graph 300, x-axis 302 is representative of a distance d from a class and y-axis 304 is representative of the confidence value. The graph 300 illustrates a sigmoid function 306 representative of the confidence function. The graph illustrates a class boundary 308 on the sigmoid function 306. Positive distance values on the x-axis 302 represent 'sepsis class' and negative distance values represent 'not sepsis class'. Higher absolute distance value represents higher confidence in the sepsis probability value. In this

example, the model threshold value (7) was 0.31 and the sigmoid function parameter (fi) value was 3.8448.
[0041] FIG. 4 illustrates a transparent tree model 400 in accordance with an exemplary embodiment. The working of a tree based sepsis model is explained herein with reference to an example case. The tree model of 400 includes a plurality of nodes 402, 406, 410, 414, 418, 424, 426, 428 a plurality of links 404, 408, 412, 416, 420, 430, 432, and a plurality of leaf nodes 434, 422. It should be noted herein that only some nodes, links and leaf nodes in the FIG. 4 are represented as required for explaining the transparent tree model 400. The node 402 is referred herein as root node and has no parent node. Each of the plurality of leaf nodes has one node among the plurality of nodes as a parent node and has no child node. Each of the other plurality of nodes has one parent node and two child nodes. The root node 402 has two child nodes 406 and 424. The node 418 has two leaf nodes 434 and 422 as child nodes. For an available set of parameters, a path in the tree model 400 may be traversed and sepsis condition for the patient may be predicted. The example case having a feature parameter set of SOFA=8, CSRU=1, RR = 23, SEP = 120 and PREV_HOSPITAL = 1 is considered. The feature parameters SOFA, RR, and SEP have a corresponding feature threshold values of 9.0, 21.31, and 111.83 respectively. The root node 402 corresponds, to the MICU (whether the patient is served by medical ICU) and since the patient is served by cardiac surgery recovery unit (CSRU), the link 404 corresponding to MICU=0 is traversed to reach the node 406 as illustrated. The node 406 corresponds to SOFA parameter and since the parameter set has a SOFA value of eight, the link 408 is traversed to reach the node 410 corresponding to RR parameter. Since the RR is equal to twenty three in the parameter set, the link 412 is traversed reaching to the node 414 corresponding to SBP parameter. The value of SBP is equal to one hundred and twenty and the link 416 is traversed to reach the node 418 corresponding to PREV_HOSPITAL parameter. Based on the PREVHOSPITAL value of one, the link 420 is traversed to reach the leaf node 422 representative of a sepsis condition with a sepsis probability value of 0.54.
[0042] The path traversed in the tree model 400 based on the parameter set is highlighted in the Fig. 4. The highlighted path through the tree of the tree model 400 from root node 402 to the leaf node 422 enhances the understanding of care giving staff about the medical

condition, and rationale behind the decision related to treatment options. The tree model 400 is amenable for visualization of intermediate steps in the prediction process in a transparent manner and is referred herein as 'transparent model'. Although the transparent models are explained with reference to tree based models, other models that are used in other embodiments are also amenable for visualization. As an example, a neural network model is considered. The neural network model has one or more input nodes, a plurality of links, one or more layers of hidden nodes and an output node. A trained neural network model has a link weight associated with each of the plurality of links. The plurality of link weights associated with the plurality of links of the neural network model may be used to highlight a path from the input node to the output node through one more of the hidden nodes.
[0043] FIG. 5 is a flow chart 500 illustrating a method for sepsis prediction for a patient in accordance with an exemplary embodiment. The method includes providing a model configuration comprising a sepsis model, definition of a feature vector, and a confidence function 502. In one embodiment, the model configuration is provided by a user or an operator. The definition of the feature vector includes a list of feature parameters, an arrangement of the feature parameters in the feature vector, and mathematical formulae required to compute the feature parameters. In one embodiment, the sepsis model and the confidence function are designed apriori for the feature vector. The model configuration may also include a technique for determining the proxy data, such as an imputation method or a surrogate substitution method. The method further includes obtaining clinical data comprising measured clinical data and stored clinical data 504. In one embodiment, the model configuration and the clinical data are stored in the memory and processed by the signal acquisition module. The measured clinical data and the stored clinical data includes a plurality of chart events such as heart rate, respiration rate Scv02 readings, and Sp02 readings, readings from a plurality of lab tests such as white blood cell count, and lactate, a plurality of patient demographic parameters such as weight, height, age and gender. Patient demographics may further include parameters such as ethnicity and family history that may give insight to the medical processing. The measured clinical data and the stored clinical data may include other parameters from a plurality of other chart events, lab tests and patient demographic data as well as other data generated within the hospital.

[0044] A plurality of feature parameters of the feature vector are estimated based on the measured clinical data 506. The plurality of feature parameters of the feature vector may be a statistical average of a plurality of readings of one of the chart events, a reading of a lab test, and a statistical average of a patient demographic parameter. The plurality of feature parameters also include standard scores such as SOFA, SAPS, and SIRS. It should be noted herein that the plurality of feature parameters may not include all the feature parameters of the feature vector and in some embodiments, at least one feature parameter of the feature vector may be missing from the plurality of feature parameters.
[0045] The proxy data corresponding to at least one missing parameter of the feature vector is determined based on the stored clinical data 508 using either the imputation method or the surrogate substitution method. In one embodiment, the proxy data is an estimate of the at least one missing parameter. As an example, the estimate of the at least one missing parameter is a mean of a plurality of readings of a ■chart event from a plurality of similar patients. Similar patients refers to patients with some relevance to the current patient. As another example, the estimate of the at least one missing parameter is a median value of a plurality of readings of a lab test from a plurality of similar patients. In another example, the estimate of the at least one missing parameter is a standard deviation of a plurality of readings of a demographic parameter corresponding to a plurality of patients. In an alternative embodiment, the proxy data corresponding to the at least missing parameter is determined by selecting a surrogate feature when the surrogate substitution method is employed.
[0046] A sepsis probability value may be determined based on the sepsis model and the feature vector 510. The sepsis model used to generate the sepsis probability value includes, but not limited to, one of a tree based model, a neural network, and a regression model. A confidence value is determined based on the sepsis probability value and the confidence function 512. In one embodiment, the confidence function is a sigmoid function. In another embodiment, the confidence function is a linear function with a suitable slope value. In one exemplary embodiment, determining the confidence value includes determining a statistical significance on a plurality of leaf nodes in a tree model.

[0047] The confidence value generated based on the confidence function and the sepsis probability value is compared with a confidence threshold value 514. In certain embodiments, the confidence threshold value is provided by an operator or is stored in the memory. The confidence value higher than the confidence threshold value indicates that the sepsis probability value is reliable. A sepsis condition is predicted during present treatment and the sepsis probability value, and the confidence value 516 are provided. A sepsis condition prediction is determined based on the sepsis probability value and a model threshold value. In some embodiments, the model threshold value is provided by an operator, or is computed apriori and stored in the memory. The confidence value indicates reliability of the sepsis condition prediction. Upon a prediction for a sepsis condition for a particular patient, appropriate remedial measures can be taken to lower the likelihood of the sepsis condition and mitigate the sepsis. For example, the patient can be given antibiotics, subject to greater medical attention, isolated from other patients, and generally treat the patient with the informed knowledge that sepsis is likely.
[0048] A confidence value lesser than the confidence threshold value indicates that the sepsis probability value 510 is not reliable. Exemplary embodiments disclosed herein provide an opportunity to improve the reliability of the sepsis probability value. In one embodiment, the measured clinical data required to identify the at least one missing parameter is determined 518 and presented to an operator or a care giving staff. The care giving staff may proceed to conduct additional laboratory tests and acquire the measured clinical data required 520 to determine the at least one missing parameter. With the availability of additional feature parameters of the feature vector, the sepsis prediction having a higher confidence value may be determined. It should be noted herein that steps of generating the measured clinical data required 520 and re-computing the sepsis probability value and the confidence value are optional and needs intervention of the care giving staff. However, the disclosed method enable the care giving staff to effectively intervene to improve the confidence value associated with the sepsis prediction value by identifying measured clinical data required to determine the at least one missing feature parameter.
[0049] Exemplary embodiments disclosed herein provide an automatic identification of hospitalized patients having a risk of developing sepsis. The disclosed technique is capable

of generating a prediction of sepsis condition based on the available parameters of the feature vector. The sepsis prediction is provided with a confidence value-and an indication of the measured clinical data required for enhancing the confidence value. The proposed technique is also capable of providing a visual representation of intermediate steps in the sepsis prediction. The visual representation indicates the feature parameters of the feature vector that influenced the sepsis prediction enhancing the transparency of the prediction process to the patient and'to the care giving staff.
[0050] It is to be understood that not necessarily all such objects or advantages described above may be achieved in accordance with any particular embodiment. Thus, for example, those skilled in the art will recognize that the systems and techniques described herein may be embodied or carried out in a manner that achieves or improves one advantage or group of advantages as taught herein without necessarily achieving other objects or advantages as may be taught or suggested herein.
[0051] While the technology has been described in detail in connection with only a limited number of embodiments, it should be readily understood that the specification is no; limited to such disclosed embodiments. Rather, the technology can be modified to incorporate any number of variations, alterations, substitutions or equivalent arrangements not heretofore described, but which are commensurate with the spirit and scope of the claims. Additionally, while various embodiments of the technology have been described, it is to be understood that aspects of the specification may include only some of the described embodiments. Accordingly, the specification is not to be seen as limited by the foregoing description, but is only limited by the scope of the appended claims.

CLAIMS:
1. A computer implemented method, comprising:
providing a model configuration comprising definition of a feature vector, a sepsis model, and a confidence function;
obtaining clinical data about a patient, wherein the clinical data comprises stored clinical data and measured clinical data;
estimating on a processor, a plurality of feature parameters of the feature vector based on the measured clinical data;
determining proxy data based on the stored clinical data, corresponding to at least one missing parameter of the feature vector, if the at least one missing parameter is different from the plurality of feature parameters;
predicting on the processor, a sepsis probability value based on the sepsis model, the feature vector, and the proxy data;
generating on the processor, a confidence value based on the sepsis probability value and the confidence function;
identifying the measured clinical data required to determine the at least one missing parameter, if the confidence value is less than a confidence threshold value; and
generating a sepsis condition prediction based on the sepsis probability value, with the confidence value indicative of reliability of the sepsis condition prediction.
2. The method of claim 1, wherein the estimating further comprises;
estimating one or more model parameters of the sepsis model based on the stored clinical data; and

. estimating one or more function parameters of the confidence function based on the stored clinical data..
3. The method of claim 1, wherein determining proxy data comprises using an imputation method or a surrogate substitution method.
4. The method of claim 1, wherein the stored clinical data and the measured clinical data comprises a plurality of chart events, a plurality of patient monitor waveforms, a plurality of lab tests, and a plurality of patient demographic parameters.
5. The method of claim 1, wherein determining the plurality of feature parameters comprises determining a plurality of statistical parameters based on at least one of a plurality of heart rate values, a plurality of respiration rate values, a plurality of blood pressure values, a plurality of oxygen saturation values, and a plurality of body temperature
values.
6. The method of claim 1, wherein the determining proxy data comprises
determining a statistical estimate of the at least one missing parameter based on the stored clinical data.
7. The method of claim 1, wherein the determining proxy data comprises
selecting a surrogate feature corresponding to the at least one missing parameter.

8. The method of claim 1, wherein the sepsis model is a tree based model or a neural network.
9. The method of claim 1, wherein the confidence function comprises a sigmoid function.
. 10. The method of claim 1, wherein the identifying comprises generating a visual information based on the plurality of feature parameters, the at least one missing parameter, the sepsis probability value and the confidence value.
11. A system, comprising:
at least one processor and a memory coupled to communications bus;
a signal acquisition module stored in the memory and executable by the at least one
processor configured to:
provide a model configuration comprising definition of a feature vector, a sepsis model, and a confidence function;
obtain clinical data of a patient, wherein the clinical data comprises stored clinical data and measured clinical data;
estimate a plurality of feature parameters of the feature vector based on the measured clinical data; and
determine proxy data corresponding to at least one missing parameter of the feature vector, based on the stored clinical data, if the at least one missing parameter is different from the plurality of feature parameters;

a sepsis prediction module stored in the memory and executable by the at least one processor, the sepsis prediction module communicatively coupled to the signal acquisition module and configured to:
generate a sepsis probability value based on the sepsis model, the feature vector, and the proxy data; and
generate a sepsis condition prediction based on the sepsis probability value; and
a confidence generator module stored in the memory and executable by the at least one processor, the confidence generator module communicatively coupled to the sepsis prediction module and the signal acquisition module configured to:
generate a confidence value based on the sepsis probability value; and
identify the measured clinical data required to determine the at least one missing parameter, if the confidence value is less than a confidence threshold value;
wherein the confidence value is indicative of reliability of the sepsis condition prediction.
12. The system of claim 11, wherein the signal acquisition module is further configured to receive the model configuration comprising an imputation method and determine the proxy data corresponding to the at least one missing parameter based on the imputation method.
13. The system of claim 11, wherein the signal acquisition module is configured to obtain a plurality of chart events, a plurality of patient monitor waveforms, a plurality of lab tests, and a plurality of demographic parameters corresponding to the patient.

14. The system of claim 11, wherein the signal acquisition module is configured to determine the proxy data corresponding to the at least one missing parameter by selecting a surrogate feature.
15. The system of claim II, wherein the sepsis model comprises a tree based
model or a neural network model.
16. The system of claim 11, wherein the confidence generator module is further configured to generate the confidence value based on a sigmoid function.
17. The system of claim 11, wherein the confidence generator module is configured to generate a visual information based on the plurality of feature parameters, the at least one missing parameter, the sepsis probability value and the confidence value.
18. A non-transitory computer readable medium encoded with a program to instruct at least one processor to:
provide a model configuration comprising definition of a feature vector, a sepsis model, and a confidence function;
obtain clinical data of a patient, wherein the clinical data comprises stored clinical data and measured clinical data;
determine a plurality of feature parameters of the feature vector based on the measured clinical data;

estimate proxy data corresponding to at least one missing parameter of the feature vector, based on the stored clinical data, if the at least one missing parameter is different from the plurality of feature parameters;
predict a sepsis probability value based on the sepsis model, and the feature vector, and the proxy data;
generate a confidence value based on the sepsis probability value and the confidence function;
identify the measured clinical data required to determine the at least one missing parameter, based on the confidence value, if the confidence value is less than a confidence threshold value; and
generate a sepsis condition prediction based on the sepsis probability value, with the confidence value indicative of reliability of the sepsis condition prediction.