TRENDS AND BASELINE DEVIATION FOR CONICAL DECISION SUPPORT

TECHNICAL FIELD

[0001] This disclosure relates generally to processing information, and more specifically, but not exclusively, to a model-based approach for determining or predicting the condition of a patient.

BACKGROUND

[0002] Clinical decision support (CDS) algorithms for early prediction of patient deterioration events (e.g., hemodynamic instability, infection, etc.) rely on clinical signatures. These signatures are based on current conditions that relate to the patient across a variety of physiological data. In some cases, the clinical signatures include a complex mixture of present absolute values (e.g., current heart rate) and deviations from a previous baseline (e.g., current heart rate relative to a previous healthy baseline). However, current approaches are inadequate because of their inability to formulate a baseline for a patient with little or no prior patient history.

SUMMARY

[0003] A brief summary of various example embodiments is presented below. Some simplifications and omissions may be made in the following summary, which is intended to highlight and introduce some aspects of the various example embodiments, but not to limit the scope of the invention. Detailed descriptions of example embodiments adequate to allow those of ordinary skill in the art to make and use the inventive concepts will follow in later sections.

[0004] In accordance with one embodiment, a method for performing health management includes receiving physiological signals for a patient; deriving one or more macro trends from the physiological signals; determining deviations of the one or more macro trends from control data; and generating a classifier based on the deviations, wherein the classifier is configured to generate a prediction for a condition of the patient. The prediction may be a health condition that the patient is expected to develop in the future.

[0005] Determining the deviations may include generating a variogram including a first curve corresponding to the one or more macro trends and a second curve corresponding to the control data, wherein the deviations of the one or more macro trends from the control data corresponds to a deviation between the first curve and the second curve over a range of time lags.

[0006] Generating the classifier may include generating a feature that summarizes the deviation over the range of time lags. Generating the feature may include generating the feature based on a weighting function. The weighting function may be a linear discriminant function. Deriving the one or more macro trends may include generating a statistical measure of at least one type of physiological signal of the physiological signals. The prediction may includes an importance score indicating a probability that the condition of the patient will worsen or appear in a predetermined future time period.

[0007] In accordance with one or more other embodiments, a system for performing health management includes a memory area configured to store instructions and a processor configured to execute the instructions to: receive physiological signals for a patient, derive one or more macro trends from the physiological signals, determine deviations of the one or more macro trends from control data, and generate a classifier based on the deviations, wherein the classifier is configured to generate a prediction for a condition of the patient. The prediction may be a health condition that the patient is expected to develop in the future.

[0008] The processor may be configured to determine the deviations by generating a variogram including a first curve corresponding to the one or more macro trends and a second curve corresponding to the control data, where the deviations of the one or more macro trends from the control data corresponds to a deviation between the first and second curves over a range of time lags. [0009] The processor may be configured to generate the classifier by generating a feature that summarizes the deviation over the range of time lags. The processor may be configured to generate the feature based on a weighting function. The weighting function may be a linear discriminant function. The processor may be configured to derive the one or more macro trends by generating a statistical measure of at least one type of physiological signal of the physiological signals. The prediction may include an importance score indicating a probability that the condition of the patient will worsen or appear in a predetermined future time period.

BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The accompanying figures, where like reference numerals refer to identical or functionally similar elements throughout the separate views, together with the detailed description below, are incorporated in and form part of the specification, and serve to illustrate example embodiments of concepts found in the claims and explain various principles and advantages of those embodiments. [0011] These and other more detailed and specific features are more fully disclosed in the following specification, reference being had to the accompanying drawings, in which:

[0012] FIG. 1 illustrates an embodiment of a method for predicting a patient condition;

[0013] FIG. 2 illustrates an example of time windows that may be used for the method;

[0014] FIG. 3 illustrates an example of a variogram generated by the method;

[0015] FIG. 4 illustrates an example of a summary feature generated based on the variogram; and

[0016] FIG. 5 illustrates an embodiment of a system for predicting a patient condition.

DETAILED DESCRIPTION

[0017] It should be understood that the figures are merely schematic and are not drawn to scale. It should also be understood that the same reference numerals are used throughout the figures to indicate the same or similar parts. [0018] The descriptions and drawings illustrate the principles of various example embodiments. It will thus be appreciated that those skilled in the art will be able to devise various arrangements that, although not explicitly described or shown herein, embody the principles of the invention and are included within its scope. Furthermore, all examples recited herein are principally intended expressly to be for pedagogical purposes to aid the reader in understanding the principles of the invention and the concepts contributed by the inventor(s) to furthering the art and are to be construed as being without limitation to such specifically recited examples and conditions. Additionally, the term, “or,” as used herein, refers to a non-exclusive or (i.e., and/ or), unless otherwise indicated (e.g., “or else” or “or in the alternative”). Also, the various example embodiments described herein are not necessarily mutually exclusive, as some example embodiments can be combined with one or more other example embodiments to form new example embodiments. Descriptors such as “first,” “second,” “third,” etc., are not meant to limit the order of elements discussed, are used to distinguish one element from the next, and are generally interchangeable. Values such as maximum or minimum may be predetermined and set to different values based on the application.

[0019] Example embodiments include a system and method for determining and/or predicting a patient condition at times including during early stages of its development or before that condition occurs. The system and method may be implemented in a variety of ways, especially, but not exclusively, when little or no baseline data for the patient is available. Such a situation may occur in situations including, but not limited to, when a new patient comes to a hospital for treatment and the hospital has little or no access to the medical history of that patient.

[0020] In accordance with one embodiment, the system and method generates a surrogate for a patient baseline. The surrogate may be generated by calculating statistical deviations from control information based on one or more physiological signals captured over a variety of time lags. These deviations (which may be referred to as “macro trends”) may be used as a basis for predicting existing or emerging changes in the state of the patient over a future time period. The period of time may be a predetermined time set, for example, based on the range of relevance of the data used to generate a corresponding prediction model. In one non-limiting example, the predetermined future time period may be relatively long (e.g., 24 to 48 hours), thus giving medical personnel sufficiently advanced notice to prevent, treat, or otherwise manage the developing adverse condition.

[0021] The method may include operations that generate and train the prediction model based on the macro trends. In one case, the model may generate a feature importance score that provides an indication of the severity and/ or probability of the adverse condition actually forming in the patient or worsening when the condition has already manifested in its early stages. An example of generating an importance score combines information from current values and baseline deviations.

[0022] FIG. 1 illustrates an embodiment of a method for determining and/ or predicting a patient condition, preferably but not exclusively during its early stages or before the condition develops. The method may be performed for a variety of patients, including ones for which little or no medical history is known. In other cases, the patient may have an accessible medical history, but not for the current problem which the patient is experiencing. The patient condition may be an adverse condition resulting from a current medical problem, a recently performed medical procedure or another type of health-related condition. Examples of adverse conditions include developing sepsis or another type of bacterial or viral infection, the occurrence of a side effect from treatment or medication, and relapse of a disease or other condition.

[0023] The method may be implemented in at least two stages. The first stage includes calculating surrogate information based on deviations from control data. The second stage includes integrating the surrogate information into a clinical decision support (CDS) model that generates information that may be used as a basis for determining the patient condition and/ or for predicting a condition of the patient that may likely develop at a future point in time. In one implementation, the model may be formulated based on one or more algorithms that predict patient deterioration in such a way that interpretability is maintained. The interpretability may be maintained, for example, by estimating one or more feature importance scores for individual physiological signals (e.g., heart rate). The scores may be generated in a way that combines importance relevant features from current information and deviations from baseline (derived from the surrogate signal in the first phase).

[0024] Referring to FIG. 1, the method includes, at 110, developing one or more surrogate baselines for the patient of interest. In one embodiment, each surrogate baseline may be developed based on one or more types of physiological signals received from at least one patient monitoring system. The physiological signals may be captured and stored in a database over a predetermined time window. Examples of the physiological signals include vital sign data and laboratory test data. The time window may be a predetermined fixed or adjustable observation period during which the physiological signals are collected. The observation period may be long enough to generate a sufficient amount of data to attain an accurate surrogate baseline for the patient. This amount of time may vary, for example, based on the condition of the patient and/ or the type of adverse conditions to be predicted. When multiple surrogate baselines are generated, each of the surrogate baselines may correspond to a different physiological parameter or condition of the patient.

[0025] At 120, data derived from the physiological signals may be processed to estimate one or more macro trend features. These macro trends may then be used as a basis for developing a patient prediction model. The macro trends may be developed based on the data derived from one type of physiological signal or data derived from a combination of types of physiological signals. The physiological signals may be indicative of health parameters, including but not limited to one or more types of vital signals, laboratory tests, or other information. The macro trends may provide an indication of the deviation of one or more of the parameters to control information. [0026] In one embodiment, the macro trends may be estimated by determining a statistical basis for each type of physiological signal collected. The statistical basis may take the form of a curve to be included in a variogram generated for each of the one or more signals. The curve may be generated over a predetermined time window. Examples of how the curve to be included in the variogram may be generated are discussed in greater detail below. In one embodiment, the statistical basis determined for the physiological signals may be different from a variogram.

[0027] At 130, deviations of the macro trend(s) are determined relative to control information. When the macro trend(s) are generated as a statistical curve to be included in a variogram, then a curve for the control information may be generated and included in the variogram relative to the statistical curve for the macro trend(s). The control information may be generated from a control group of patients that have exhibited similar physiological conditions as the patent under observation. This information may be generated at some time point in the past based on verified data, that establishes the credibility of the control information for given types of physiological signals. By plotting the data values of the control information over predetermined time periods, a statistical curve may be generated and included in the variogram relative to the curve for the macro trend(s) of the patient.

[0028] Each generated variogram may therefore identify and track deviations of the macro trend curve(s) from the control information curve regarding one or more physiological aspects or conditions of the patient. In one embodiment, each of the variograms may specify the deviations over one or more predetermined time lags. While the generation of these variograms may be considered beneficial for some applications, the macro trends may be expressed in a variety of other ways including, but not limited to, raw data, standard deviations, pattern matching, and threshold or signature comparisons, as well as through other methods.

[0029] At 140, the macro trends and/or their deviations from the control information may be processed to generate and train a classifier that may be used to predict the present and/ or future condition of the patient. The classifier may be implemented as a model that is generated and trained, for example, machine -learning techniques. In one embodiment, the model may include a decision tree formed based on the macro trends developed for the patient. The model may take a different form in another embodiment.

[0030] Different processing techniques may be used to form the model. For example, the variogram- based macro trends may be averaged over one or more of a time window and then assigned weights offset by time lags. The weights may be assigned, for example, based on an integral-based weighting function designed to provide an indication of the severity of the deviations of the patient macro trends versus the baseline (e.g., control information) . In one example, the weights may provide an indication of the degree of deviation from the baseline of one feature relative to one or more other features. (The weights may correspond to the g functions discussed in greater detail below.) In addition to these features, the time component of the classifier may be used to provide a time scale as to the prediction, including but not limited to when the prediction of an adverse condition of the patient may form in the coming hours or days.

[0031] At 150, the classifier may be applied to generate information including a prediction of the condition of the patient. As previously indicated, the condition may be a present condition of the patient (e.g., which may be at the onset of early stages of an infection or other adverse condition) and/ or may be a future condition that likely will develop. Thus, through use of the classifier, medical personnel may determine the likely outcome of the condition of the patient, for example, before it occurs or at the early stages, even in situations when little or no medical history is known about the patient. Such a classifier may therefore give medical professionals forewarning and lead them to take corrective action to preempt the formation or further progression of a deterioration in health. Macro Trend Deviation

[0032] In one embodiment, deviation of the curves corresponding to the macro trends and control data may be derived from physiological signals that discriminate between two cohorts of patients. In one example of clinical decision support (CDS) algorithm, the two cohorts of patients may include a first cohort that corresponds to a deteriorating group of patients, for example, with respect to a particular organ system and a second cohort that corresponds to a control group. Each set of macro trend features may be derived for a particular physiological signal (e.g., heart rate). When multiple macro trend sets are to be derived, each set may correspond to one or more different physiological signals, laboratory test parameters/ results, and/ or other heath conditions or measurements.

[0033] In one embodiment, an initial operation for determining macro trends includes defining a training dataset. This may be mathematically explained as follows. Tet X;(t ) denote a physiological time-series (e.g., heart rate, temperature, respiration rate) for the i* patient during an observation window that precedes a forecast window. During the observation window, the patient may experience an acute deterioration event, e.g., hemodynamic instability or infection. If the patient is a control patient, the forecast window may be a random time selected during the hospital stay of the patient. [0034] FIG. 2 illustrates an example of the observation window 210 and the forecast window 220 arranged relative to a timeline 201. During the observation window, one or more types of physiological data signals are collected for the patient. The physiological data may include, for example, one or more types of vital signs 230 (e.g., heart rate, temperature, respiration rate) and/ or data or results derived from one or more laboratory tests 240. The forecast window 220 may occur later in time from the observation window and may be based on one or more target variables, which provide an indication of predicted patient deterioration, the time for such a deterioration to develop, and/or how the deterioration may be prevented, mitigated, or otherwise controlled or managed. [0035] In addition to X;(t ), a binary label yi may be defined that indicates if the patient experienced a deterioration event or was a control patient. In one embodiment, may have a first value (e.g., 1) if a deterioration event occurred during the observation window 210 and a second value (e.g., 0) if the patient is a control patient. The duration of the observation window 210 may be a predetermined time period, e.g., 48 hours. The duration of the observation window may be treated as a hyperparameter set to an initial value. The duration may then be adjusted during the course of training or machine learning for the model. The duration of the observation window may lie within a range of values determiend, for example, based on practical factors to be considered, e.g., typical hospital length of stay before an acute deterioration event tends to develop, etc.

[0036] In one embodiment, each macro trend for a patient may be calculated (or estimated) as a curve to be included in a variogram genearted for at least one physiological signal. The variogram may provide an indication of the average deviation in the physiological signal as a function of time lag, or stated differently between the curve generated based on the physiological signals of the patient taken relative to a curve of data belonging to a control cohort. In some cases, the average deviation may change for different time lags in the variogram.

[0037] Each variogram E _j (t) may be calculated, for example, based on Equation (1): where E is an estimation function, Xi represents physiological signal i (e.g., blood pressure, heart rate, etc.), ti and t2 are sampling times for the physiological signal during the observation window, and h corresponds to time lag between times.

[0038] More specifically, using Equation (1) a variogram may estimate the expected mean absolute deviation in the physiological signal ¾ for time lag h. Other deviation estimates are possible (e.g., mean square error, etc.), although mean absolute deviation may provide robustness to noise for at least some applications. Due to the discrete sampling of the data, the variogram may practically be estimated by taking all pairs of measurements (x(ti),x(t ₂ )) and binning them based on the absolute measurement time difference |E - t ₂ | . For example, let N(h) denote the set of measurement pairs binned to time lag h. Then, the corresponding variogram estimate, V,(h), corresponds to the average signal deviation | x,(ti) - Xi(t ₂ ) I over all pairs (ti,t ₂ ) in N(h).

Weight Calculations

[0039] As indicated, a variogram generated based on Equation (1) may indicate an average deviation calcualted as a function of time lag for each patient. In a next operation, a single feature may be determined that summarizes the deviation over all time lags. This may be performed in a variety of ways. One way is to employ a weighting scheme that maximizes discriminative information between the two cohorts of patients, e.g., the control cohort and the deterioration cohort. The weighting scheme may be implemented as follows.

[0040] Initially, a summary macro trend feature fi may be derived based on Equation (2).

[0041] Equation (2) computes a weighted average of the variogram based on a weighting function w(h). The weighting function may be derived in a number of ways. For example, initially, a training dataset may be generated with a feature space that corresponds to variogram estimates at d distinct time lags [Vi(hi),Vi(h2),... ,Vi(h _d )], where d is an integer and corresponding label yi corresponds to the same deterioration/ control label as previuosly defined. Then, a classifier may be trained to predict the label yi from the d variogram time lags. The weighting function is then given by the classifier weights. (In some embodiments, the term “time lag” may be synonymous with time points relative to a predetermined reference time, e.g., the time the patient was admitted to a hospital, began to be treated, when the patient was first tested and/ or vital signs taken, and/ or another time) .

[0042] In one embodiment, the weighting function may be a linear discriminant function. In such a case, let V^(h) and V^(h) denote the average variograms for the deterioration group (y; = 1) and control group ( ; = 0), respectively. The variogram V^(h) may be calculated for the deterioration group based on Equation (3) and variogram V^(h) may be calculated for the control group based on Equation (4). where Ni and N ₀ denote the number of deterioration and control patients, respectively, and the summation equations å denote population covariance of the variogram over time lags. An example of the linear discriminant weighting function is given by Equation (5) .

[0043] Figure 3 illustrates a variogram generated for the example of the physiological signal type heart rate. In FIG. 3, statistical curve 310 is an averaged heart rate variogram for a deterioration group, which in this case includes patients with early signs of infection. Statistical curve 320 is an averaged heart rate variogram for patients in a control group. From the curves in this variogram, it is evident that the deviation between the two cohorts of patient groups increases for longer time lags. For example, as illustrated in FIG. 3, a first deviation D1 may exist at a 24-hour time lag, a second deviation D2 may exist at a 36-hour time lag, and a third deviation D3 may exist at a 42-hour time lag. This indicates that a discriminating signal in the variogram arises from macro trends that compare patient state on longer time scales (e.g., 24+ hours ago). In the absence of baseline information for a specific patient (who, for example, has just come into a hopsital), this surrogate signal (e.g., variogram curve 310 or the difference between the two variogram curves 310 and 320) may be used to identify emergent changes in the physiological state of the patient.

[0044] FIG. 4 illustrates an example of a curve 410 generated by a weighting function for the heart rate variogram in FIG. 3. The weighting function curve is generated based on the linear discriminant method and associated equations, as described above. As shown, curve 410 increases for longer time lags, thereby confirming the deviation between statistical curve 310 and control group curve 320. In some cases, the linear discriminant method may generate more discriminant results than other trend estimation techniques, e.g., based on the slope of a linear trend line. This may be due to a variety of factors. For example, the variogram looks at overall changes from earlier time periods and does not assume directionality (e.g., increasing or decreasing trend). Also, the variogram is less sensitive to the trajectory/ evolution of the physiological time-series, whereas trend lines may assume a linearly increasing/ decreasing trend in the underlying physiological signal. Also, the variogram method may model macro trend changes in the signal more broadly.

[0045] The aforementioned operations may be repeated for multiple physiological signals. This may result in the formation of an augmented feature space for training a predictive model. Such a predictive model may be generated based on an algorithm that integrates the macro trend features in an interpretable way, an example of which is described below.

[0046] For each patient i in a training dataset, let denote a state vector for the patient. The state vector may include the last-measured value x ⁿ⁾ and macro trend feature _// ⁿ⁾ for each of d physiological signals, n=l,...,d. To develop an interpretable model that can predict the corresponding label >v from the set of features, machine-learning techniques may be used to learn a generalized additive model, for example, as expressed in Equation (6).

[0047] In Equation (6), features from the underlying physiological signals are grouped together so that g _n are nonlinear functions that operate only on the features derived from the nth physiological signal. Thus, nonlinear functions g _n may be used as a basis to predict deterioration of the patient based on the current value (s) of the physiological signal (s) (x ⁿ⁾ ), along with the macro trend features derived from that signal (/? ⁿ⁾ ). The nonlinear functions g _n may be learned in a variety of ways. For example, one way involves a technique which involves boosting with decision trees. When this technique is implemented, on each boosting round, a physiological signal may be selected and then the decision tree may be developed using only features derived from that signal.

[0048] By constraining the predictive model in the manner indicated above, an importance score may be assigned to each physiological signal. For example, the importance score for the n ^th physiological signal may be based on the magnitude of g _n ixi^ji^) relative to the outputs from the other physiological signals. This enables a dynamic feature importance visualization that ranks the physiological signals based on the state of a particular patient at point-of-care.

[0049] To increase interpretability, the time-courses of the most important features may be visualized along with empirical variogram curves Vi(h) contrasted against the average variogram V^(h), as previously described, for the control population.

[0050] FIG. 5 illustrates an embodiment of a system which may be used to implement all or a portion of the operations of the method embodiments described herein. The system includes a processor 510, a memory 520, a database or other type of data storage area 530, an interface 540, and an output device 550. The processor 510 executes instruction stored in the memory 520 in order to perform operations of the method embodiments. As such, the memory 520 may be a non-transitory computer-readable medium and the instructions may store one or more algorithms for predicting the medical condition of a patent based on the models described herein. Data for training the model may be stored in the data storage area 530, along with other forms of patient and system data for performing the predictions. Once the model is generated and trained, the interface 540 may receive information about a specific patient that maybe input into the model to generate predictions. The output device 550 may be a mobile device, computer, workstation, tablet, or any other device capable of interacting with the processor.

[0051] In one example, the output devicer 550 may output importance scores in graphical and/or textual format for the patient. The importance scores may be computed as previuosly described (e.g., based on the magnitude of g _n (x ⁿ⁾ ,// ⁿ⁾ )). If the importance score for a particular feature (e.g., heart rate) is elevated, then, in one embodiment, the interface may visualize the patient’s variogram trajectory along with the control variogram.

[0052] The method and system embodiments described herein may be performed by code or instructions to be executed by a computer, processor, controller, or other signal processing device. The code or instructions may be stored in a non-transitory computer-readable medium in accordance with one or more embodiments. Because the algorithms that form the basis of the methods (or operations of the computer, processor, controller, or other signal processing device) are described in detail, the code or instructions for implementing the operations of the method embodiments may transform the computer, processor, controller, or other signal processing device into a special-purpose processor for performing the methods herein.

[0053] The processors, algorithms, models, and other signal generating and signal procesing features of the embodiments disclosed herein may be implemented in logic which, for example, may include hardware, software, or both. When implemented at least partially in hardware, the processors, algorithms, models, and other signal generating and signal procesing features may be, for example, any one of a variety of integrated circuits including but not limited to an application-specific integrated circuit, a field-programmable gate array, a combination of logic gates, a system-on-chip, a microprocessor, or another type of processing or control circuit.

[0054] When implemented in at least partially in software, the processors, algorithms, models, and other signal generating and signal procesing features may include, for example, a memory or other storage device for storing code or instructions to be executed, for example, by a computer, processor, microprocessor, controller, or other signal processing device. Because the algorithms that form the basis of the methods (or operations of the computer, processor, microprocessor, controller, or other signal processing device) are described in detail, the code or instructions for implementing the operations of the method embodiments may transform the computer, processor, controller, or other signal processing device into a special-purpose processor for performing the methods herein.

[0055] It should be apparent from the foregoing description that various exemplary embodiments of the invention may be implemented in hardware. Furthermore, various exemplary embodiments may be implemented as instructions stored on a non-transitory machine-readable storage medium, such as a volatile or non-volatile memory, which may be read and executed by at least one processor to perform the operations described in detail herein. A non-transitory machine-readable storage medium may include any mechanism for storing information in a form readable by a machine, such as a personal or laptop computer, a server, or other computing device. Thus, a non-transitory machine- readable storage medium may include read-only memory (ROM), random-access memory (RAM), magnetic disk storage media, optical storage media, flash-memory devices, and similar storage media and excludes transitory signals. [0056] Although the various exemplary embodiments have been described in detail with particular reference to certain exemplary aspects thereof, it should be understood that the invention is capable of other example embodiments and its details are capable of modifications in various obvious respects. As is readily apparent to those skilled in the art, variations and modifications can be affected while remaining within the spirit and scope of the invention. Accordingly, the foregoing disclosure, description, and figures are for illustrative purposes only and do not in any way limit the invention, which is defined only by the claims.