Predictive risk model optimization

Abstract

The present application relates to predictive risk model optimization. The system disclosed herein includes a hardware processor and predictive risk model training software code stored in system memory. The hardware processor executes the software code to receive vital sign data regarding a health state for a population of subjects including positive and negative subjects, define a data set for training a predictive risk model, convert the vital sign data into parameters characterizing the vital sign data, and obtain differential parameters based on the parameters. The hardware processor executes the software code to further generate combined parameters using the parameters and the differential parameters, analyze the parameters, the differential parameters, and the combined parameters to identify a reduced set of parameters correlated to the health state, identify a predictive set of parameters that enable prediction of the health state of a living subject, and calculate predictive risk model coefficients.

Description

[0001] This application is a divisional application of Chinese patent application 2017800342227 (PCT / US2017 / 042363) entitled "Optimization of Predictive Risk Model" filed on July 17, 2017. Background Technology

[0002] Critically ill patients and those undergoing surgery are at risk of developing a number of serious physiological conditions that, if not detected and treated promptly and effectively, can lead to irreversible organ damage or even death. Examples of these conditions include hypotension, hypovolemia, acute blood loss, septic shock, and cardiovascular failure or “collapse,” to name just a few. For each of these conditions, early detection is crucial to prevent permanent damage to the patient. Even better is the ability to predict the onset of some or all of these conditions in advance, allowing for the preparation of appropriate medical interventions.

[0003] As a concrete example, hypotension, or hypotensive disorder, can be a precursor to serious medical complications in patients undergoing surgery and critically ill patients treated in the intensive care unit (ICU). In the operating room (OR) setting, intraoperative hypotension is associated with increased mortality and organ damage. Furthermore, hypotension is relatively common and is often considered one of the initial signs of deterioration in patients in the OR and ICU. For example, hypotension occurs overall in up to approximately 33% of surgeries and in up to 85% of high-risk surgeries. In ICU patients, hypotension occurs in approximately 24% to approximately 85% of all patients, with an 85% incidence in critically ill patients.

[0004] Traditional patient monitoring for the aforementioned hypotension and other serious physiological conditions may involve continuous or periodic measurements of vital signs such as blood pressure, pulse rate, and respiration. However, whether performed continuously or periodically, this monitoring typically only provides a real-time assessment of the patient's condition. Therefore, these conditions are often only detected after the onset of hypotension and other serious physiological conditions, leading to a delay in remedial measures and interventions until the patient's condition has already begun to deteriorate. However, these physiological conditions can quickly lead to potentially devastating medical consequences. For example, even relatively mild hypotension can predict or cause cardiac arrest in patients with limited cardiac reserve.

[0005] Given the susceptibility of OR and ICU patients to hypotension and other potentially dangerous physiological states, and the serious and sometimes immediate medical consequences that may result when patients enter these states, there is a great need for solutions to predict future patient deterioration caused by conditions such as hypotension, hypovolemia, acute blood loss, or collapse. Summary of the Invention

[0006] Systems and methods for optimizing predictive risk models are provided, substantially as shown and / or described in conjunction with at least one figure, and set forth more fully in the claims. Attached Figure Description

[0007] Figure 1 A diagram of an exemplary system for training a risk prediction model according to one embodiment of the present disclosure is shown;

[0008] Figure 2 Another exemplary implementation of a system for training a predictive risk model is shown;

[0009] Figure 3 An exemplary system according to one embodiment is illustrated, and a computer-readable non-transitory medium including instructions for performing training of a predictive risk model is shown; and

[0010] Figure 4 This is a flowchart illustrating an exemplary method used by a system to train a predictive risk model according to one implementation. Detailed Implementation

[0011] The following description contains specific information relating to embodiments described in this disclosure. Those skilled in the art will recognize that this disclosure may be practiced in ways other than those specifically discussed herein. The accompanying drawings and detailed descriptions in this application are for exemplary embodiments only. Unless otherwise stated, the same or corresponding elements in the drawings may be indicated by the same or corresponding reference numerals. Furthermore, the drawings and illustrations in this application are generally not to scale and are not intended to correspond to actual relative dimensions.

[0012] This application relates to systems and methods for providing improved patient care through predictive risk modeling of various potentially dangerous physiological states that critically ill or surgical patients may be susceptible to. To this end, this application discloses systems and methods for training predictive risk models to significantly optimize the reliability of the model in predicting the occurrence of such dangerous physiological states in patients. Examples of these dangerous physiological states include hypotension, hypovolemia, acute blood loss, septic shock, and cardiovascular failure or “collapse,” to name just a few. According to various embodiments, the systems and methods disclosed in this application can be used by healthcare professionals to predict dangerous physiological states before their onset. Due to this early warning, the systems and methods disclosed in this application enable the preparation of effective medical interventions for early treatment of or complete prevention of the anticipated condition.

[0013] Figure 1 A diagram of an exemplary system for training a risk prediction model according to one embodiment is shown. Figure 1As shown, system 102 is located within communication environment 100, which includes communication network 120, client system 130, system users 140, positive subject group 150, and negative subject group 154.

[0014] System 102 includes a hardware processor 104 and a system memory 106 storing software code 110 for training the predictive risk model. Furthermore, system memory 106 is shown to include a predictive risk model 112, which includes a set of parameter predictions 114. Figure 1 The diagram also shows a network communication link 122 that interactively connects client system 130 and system 102 via communication network 120, and vital sign data 160 received by system 102 via communication network 120 from positive subject group 150 and negative subject group 154.

[0015] according to Figure 1 In the implementation shown, system user 140 (who may be, for example, a healthcare professional or medical researcher) can interact with system 102 via communication network 120 using client system 130. For example, system user 140 can receive a predictive risk model 112 including a parameter prediction set 114 via communication network 120, and / or can download the predictive risk model training software code 110 to client system 130 via communication network 120. In one implementation, system 102 may correspond to one or more web servers, for example, which can be accessed via a packet network such as the Internet. Alternatively, system 102 may correspond to one or more servers supporting a local area network (LAN) or included in another type of limited distribution network.

[0016] Hardware processor 104 is configured to execute predictive risk model training software code 110 to receive vital sign data 160 regarding health status for each subject in a subject group comprising a positive subject group 150 and a negative subject group 154. For example, in one exemplary embodiment, health status may be hypotension, wherein the positive subject group 150 includes only subjects who have experienced hypotension, while the negative subject group 150 includes only subjects who have not experienced hypotension. In this embodiment, for example, vital sign data 160 may include arterial pressure data for each subject.

[0017] Hardware processor 104 is also configured to execute predictive risk model training software code 110 to convert vital sign data 160 into multiple parameters characterizing vital sign data 160. Furthermore, hardware processor 104 is configured to execute predictive risk model training software code 110 to obtain differential parameters based on the multiple parameters characterizing vital sign data 160, and to generate combined parameters using the multiple parameters characterizing vital sign data 160 and the differential parameters. Hardware processor 104 is also configured to execute predictive risk model training software code 110 to analyze the multiple parameters, differential parameters, and combined parameters characterizing vital sign data 160 to identify a parameter reduction set related to health status. Hardware processor 104 is also configured to execute predictive risk model training software code 110 to identify a parameter prediction set 114 from the parameter reduction set that enables the prediction of the health status of a living subject, thereby training the predictive risk model 112 to significantly optimize its predictive reliability.

[0018] In some embodiments, hardware processor 104 is configured to execute predictive risk model training software code 110 to display, for example, parameters and / or parameter prediction set 114 characterizing the predictive risk model 112 and / or vital sign data 160 to system user 140 via display features available on client system 130. In some embodiments, hardware processor 104 is configured to execute predictive risk model training software code 110 to update or otherwise modify parameter prediction set 114 based on additional vital sign data 160 received from one or more of the positive subject population database 150 and the negative subject population database 154.

[0019] Note that, although Figure 1 The risk prediction model 112 is described as residing in system memory 106, but in some implementations, the risk prediction model 112 may be copied to a non-volatile storage device. Figure 1 (Not shown in the diagram), or as described above, can be transmitted to the client system 130 via the communication network 120. It should also be noted that although the client system 130... Figure 1 The system is shown as a personal computer (PC), but this representation is provided by way of example only. In other implementations, for example, the client system 130 may be a mobile communication device, such as a smartphone or tablet.

[0020] refer to Figure 2 , Figure 2 A more detailed exemplary implementation of the client system 230 is shown, which itself can be configured to train a risk prediction model. Figure 2 The communication environment 200 includes a client system 230, which is interactively connected to system 202 via a network communication link 222. For example... Figure 2As shown, system 202 includes a hardware processor 204 and a system memory 206 that stores the software code 210a for training the risk prediction model. Figure 2 As further shown, the client system 230 includes a display 232, a client hardware processor 234, and a client system memory 236 that stores the predictive risk model training software code 210b. Figure 2 The diagram also shows a predictive risk model 212 that includes a parameter prediction set 214.

[0021] Network communication link 222 and system 202 including hardware processor 204 and system memory 206 generally correspond to Figure 1 The network communication link 122 and the system 102 including hardware processor 104 and system memory 106. Furthermore, Figure 2 The software code 210a for training the predictive risk model corresponds to... Figure 1 The predictive risk model training software code 110. In other words, the predictive risk model training software code 210a can share any of the features attributed to the corresponding predictive risk model training software code 110, as described in this application.

[0022] Client system 230 generally corresponds to Figure 1 The client system 130 in the middle. Furthermore, the prediction risk model training software code 210b corresponds to the prediction risk model training software code 110 / 210a. Therefore, the prediction risk model training software code 210b and the prediction risk model 212 including the parameter prediction set 214 can share the attribute of Figure 1 The corresponding predictive risk model training software code 110 and the features of the predictive risk model 112 including the parameter prediction set 114 shown are either as described in this application.

[0023] according to Figure 2 In the exemplary embodiment shown, the predictive risk model training software code 210b, located in the client system memory 236, has been received from the system 202 via network communication link 222. In one embodiment, for example, network communication link 222 corresponds to the transmission of the predictive risk model training software code 210b over a packet network. Once transmitted, for example by being downloaded over network communication link 222, the predictive risk model training software code 210b can be persistently stored in the client system memory 236 and can be executed locally on the client system 230 by the client hardware processor 234.

[0024] The client hardware processor 234 can be, for example, the central processing unit (CPU) of the client system 230. The role of the client hardware processor 234 is to run the operating system of the client system 230 and execute the predictive risk model training software code 210b. Figure 2 In an exemplary implementation, the user of the client system 230 (such as...) Figure 1 The system user 140 can use the prediction risk model training software code 210b on the client system 230 to identify the parameter prediction set 214, thereby training the prediction risk model 212.

[0025] Additionally, system user 140 can utilize the predictive risk model training software code 210b on client system 230 to display the predictive risk model 212 and / or the parameters and / or parameter prediction set 214 characterizing the vital signs data 160 on display 232. Display 232 can be in the form of a liquid crystal display (LCD), a light-emitting diode (LED) display, an organic light-emitting diode (OLED) display, or other suitable display screen, which performs the physical conversion of signal to light to display the predictive risk model 212 and / or the parameters and / or parameter prediction set 214 characterizing the vital signs data 160 to system user 140.

[0026] Turn now Figure 3 , Figure 3 An exemplary system according to one embodiment and a computer-readable non-transitory medium including instructions for training and significantly optimizing a predictive risk model are illustrated. System 330 includes a computer 338 having a hardware processor 334 and system memory 336, interactively linked to a display 332. Figure 2 Similar to display 232, display 332 can take the form of, for example, an LCD, LED, or OLED display, and is configured to perform a physical conversion of signals to light in order to display, for example, parameters 316 characterizing vital signs data 160. System 330, including hardware processor 334 and system memory 336, generally corresponds to... Figure 1 System 102 and client system 130 and Figure 2 Any or all of system 202 and client system 230.

[0027] Note that parameters 316 characterizing vital sign data 160 are displayed on display 332 and include features 362, 364, 366, 368, and 358 of arterial pressure waveform 360. Features 362, 364, 366, 368, and 358 included in parameter 316 correspond to the onset of the heartbeat that generates arterial pressure waveform 360, the maximum systolic pressure marking the end of the systolic rise, the presence of the dicrotic notch marking the end of the systolic fall, the diastolic phase of the heartbeat, and an exemplary slope of arterial pressure waveform 360, respectively. Further note that parameter 316, including arterial pressure waveform 360 and features 362, 364, 366, 368, and 358, may correspond to a specific situation where the health status being generated by the predictive risk model is hypotension.

[0028] In addition to features 362, 364, 366, and 368 of the arterial pressure waveform 360 itself, the behavior of the arterial pressure waveform 360 during the intervals between these features can also be used as a parameter characterizing vital sign data 160. For example, the interval between the onset of the heartbeat at feature 362 and the maximum systolic pressure at feature 364 marks the duration of the systolic rise (hereinafter referred to as "systolic rise 362-364"). The systolic fall of the arterial pressure waveform 360 is marked by the interval between the maximum systolic pressure at feature 364 and the dicrotic notch at feature 366 (hereinafter referred to as "systolic fall 364-366"). The systolic rise 362-364 and the systolic fall 364-366 together mark the entire systolic phase (hereinafter referred to as "systolic phase 362-366"), while the interval between the dicrotic notch at feature 366 and the diastolic phase at feature 368 marks the diastolic phase of the arterial pressure waveform 360 (hereinafter referred to as "diastolic phase 366-368").

[0029] Potential diagnostic interest also includes the behavior of arterial pressure waveform 360 in the intervals from the maximum systolic pressure at feature 364 to diastole at feature 368 (hereinafter referred to as "interval 364-368"), and the behavior of arterial pressure waveform 360 from the onset of heartbeat at feature 362 to diastole at feature 368 (hereinafter referred to as "heartbeat interval 362-368"). The behavior of arterial pressure waveform 360 during the following intervals can be determined, for example, by measuring the area under the curve of arterial pressure waveform 360 and the standard deviation of arterial pressure waveform 360 in each of these intervals: 1) systolic rise 362-364, 2) systolic fall 364-366, 3) systolic 362-366, 4) diastolic 366-368, 5) interval 364-368, and 6) heartbeat interval 362-368. The corresponding areas and standard deviations measured at intervals 1, 2, 3, 4, 5 and 6 (hereinafter referred to as "intervals 1-6") can be used as additional parameters characterizing the vital signs data 160.

[0030] Figure 3The document also illustrates a computer-readable non-transitory medium 318 on which predictive risk model training software code 310 is stored. As used herein, the term "computer-readable non-transitory medium" refers to any medium that provides instructions to the hardware processor 334 of the computer 338, excluding carrier waves or other transient signals. Therefore, for example, a computer-readable non-transitory medium can correspond to various types of media, such as volatile and non-volatile media. Volatile media may include dynamic memory, such as dynamic random access memory (dynamic RAM), while non-volatile memory may include optical, magnetic, or electrostatic storage devices. Common forms of computer-readable non-transitory media include, for example, optical discs, RAM, programmable read-only memory (PROM), erasable PROM (EPROM), and FLASH memory (flash memory).

[0031] according to Figure 3 In the embodiment shown, computer-readable non-transitory medium 318 provides predictive risk model training software code 310 for execution by hardware processor 334 of computer 338. When executed by hardware processor 334, predictive risk model training software code 310 instantiates a model corresponding to... Figure 1 The prediction risk model training software code 310 / 210a / 210b in / 2 is the prediction risk model training software code 110 / 210a / 210b, and is capable of performing all operations attributed to the corresponding feature by this application. For example, when executed by the hardware processor 334, the prediction risk model training software code 310 is configured to identify the parameter prediction set 114 / 214, thereby training the prediction risk model 112 / 212.

[0032] The following is for reference. Figure 4 Further descriptions are needed for training the above reference. Figure 1 , Figure 2 and Figure 3 The system of predictive risk models is discussed. Figure 4 A flowchart 470 is presented outlining an exemplary method for training a predictive risk model by a system according to one embodiment. The method outlined in flowchart 470 can be executed using predictive risk model training software code 110 / 210a / 210b / 310 executed by hardware processor 104 / 204 / 234 / 334.

[0033] Flowchart 470 begins with receiving vital sign data 160 (action 471) regarding the health status of each subject in a subject group comprising a positive subject group 150 and a negative subject group 154. The vital sign data 160 can be received via predictive risk model training software code 110 / 210a / 210b / 310 executed by hardware processors 104 / 204 / 234 / 334 of system 102 / 202 / 230 / 330. (As...) Figure 1 As shown, vital sign data 160 can be received from positive subject group 150 and negative subject group 154 ​​via communication network 120 by the predictive risk model training software code 110 / 210a / 210b / 310.

[0034] Note that, for clarity of concept, the method outlined in flowchart 470 will be described with reference to a specific embodiment in which the predictive risk model being trained for significant optimization targets a health condition of hypotension. However, it should be emphasized that, for example, the systems and methods disclosed in this application can be applied to perform predictive risk model training and significant optimization for other health conditions of interest, such as hypovolemia, acute blood loss, sepsis or septic shock, extubation failure, postoperative complications, and cardiovascular failure or collapse. Regarding the specific and exemplary case of hypotension as the health condition of interest, vital sign data 160 may take the form of hemodynamic data corresponding to arterial pressure for each subject in the positive subject group 150 and the negative subject group 154.

[0035] Flowchart 470 continues by defining the data subset (hereinafter referred to as the “training subset”) used to train the predictive risk model 112 / 212 (Action 472). The training subset can be obtained from vital sign data 160 related to health status (e.g., hypotension in this example) of the positive subject group 150 and the negative subject group 154. The positive training subset can be defined as all time periods during which the health status occurs in the positive subject group 150. The positive training subset can also be defined as all time periods during which the health status occurs, plus the time periods preceding the occurrence of the health status in the positive subject group 150 (e.g., 5 minutes, 10 minutes, or 15 minutes before the occurrence of the health status). An example of a positive training subset is all data points when a hypotension event occurs, plus all data points 5 minutes, 10 minutes, or 15 minutes before the hypotension event.

[0036] The negative training subset can be defined as all time periods in the negative subject population 154 where no healthy state occurred. The negative training subset can also be defined as all time periods in the positive subject population 150 where no healthy state occurred, wherein the time periods must be at a certain time distance from the time periods in which the healthy state occurred. An example of a negative training subset is all data points when no hypotensive event occurred, and this time period must be at least 20 minutes before (before) and after the most recent hypotensive event. The negative training subset can also be defined as all data points for the negative subject population 154.

[0037] Flowchart 470 continues to convert vital sign data 160 into parameters 316 characterizing vital sign data 160 (action 473). This conversion from vital sign data 160 to parameters 316 can be performed by the predictive risk model training software code 110 / 210a / 210b / 310, which is executed by hardware processor 104 / 204 / 234 / 334. (See above reference...) Figure 3 The parameter 316 discussed may include features of the arterial pressure waveform 360, including the start of the heartbeat that generates the arterial pressure waveform 360 362, the maximum systolic pressure marking the end of the systolic rise 364, the presence of the diabetic notch marking the end of the systolic fall 364, the diastolic phase of the heartbeat 368, and an exemplary slope 358 of the arterial pressure waveform 360.

[0038] Additionally, parameter 316 may further include the corresponding area and standard deviation measured for intervals 1-6 of the arterial pressure waveform 360, as referenced above. Figure 3 The arterial pressure waveform 360 discussed can be the central arterial pressure waveform from any subject, such as the positive subject group 150 or the negative subject group 154.

[0039] Note that slope 358 simply represents multiple slopes that can be measured at multiple locations along the arterial pressure waveform 360. Also note that parameter 316 only provides a sample of parameters that can be converted from vital signs data 160. In practice, parameter 316 may include hundreds of parameters. Examples of additional parameters that can typically be included in parameter 316 are cardiac output, cardiac index, stroke volume, stroke volume index, pulse rate, systemic vascular resistance, systemic vascular resistance index, and mean arterial pressure (MAP). Additionally, parameter 316 may include various types of parameters found to predict future hypotension. For example, parameter 316 may include any or all of baroreflex sensitivity measurements, hemodynamic complexity measurements, and frequency domain hemodynamic characteristics.

[0040] Baroreflex sensitivity measurements quantify the relationships between complementary physiological processes. For example, a decrease in blood pressure in a healthy living subject is typically compensated for by an increase in heart rate and / or peripheral resistance. Baroreflex sensitivity measurements, derived from the 360° arterial pressure waveform, correspond to the degree to which the subject generating the 360° arterial pressure waveform responds appropriately to normal physiological changes. Hemodynamic complexity measurements quantify the regularity of cardiac measurements over time, as well as entropy, i.e., the unpredictability of fluctuations in cardiac measurements. Frequency-domain hemodynamic features quantify measurements of cardiac performance as functions of frequency rather than time.

[0041] Flowchart 470 continues to obtain the difference parameters based on parameter 316 (action 474). Obtaining the difference parameters (hereinafter referred to as "difference parameters") based on parameter 316 can be performed by the predictive risk model training software code 110 / 210a / 210b / 310, which is executed by hardware processor 104 / 204 / 234 / 334. For example, the difference parameters can be obtained by determining the variation of parameter 316 relative to time, relative to frequency, or relative to other parameters in parameter 316. As a result, each parameter in parameter 316 can generate one, two, or several difference parameters.

[0042] For example, the differential parameter stroke volume variation (SVV) can be obtained based on the change in stroke volume (SV) as a function of time and / or sampling frequency. Similarly, the change in mean arterial pressure (ΔMAP) can be obtained as a differential parameter with respect to time and / or sampling frequency, etc. As another example, the change in mean arterial pressure relative to time can be obtained by subtracting the average of the mean arterial pressure over the past 5 minutes, the past 10 minutes, etc., from the current value of the mean arterial pressure. As mentioned above, there can be hundreds of parameters 316, and one or more differential parameters can be obtained from each of the parameters 316. Therefore, there can be thousands of parameters 316 and differential parameters together.

[0043] Flowchart 470 continues to generate combined parameters using parameter 316 and the difference parameters (action 475). This generation of combined parameters can be performed using the prediction risk model training software code 110 / 210a / 210b / 310 executed by the hardware processor 104 / 204 / 234 / 334. For example, combined parameters can be generated using parameter 316 and the difference parameters by generating a power combination of a subset of parameter 316 and the difference parameters. Note that, as used in this application, the term "subset of parameter 316 and the difference parameters" refers to a subset of parameters less than 316 and less than the number of difference parameters, including some of parameter 316 and / or some of the difference parameters.

[0044] As a concrete example, each parameter in the combined parameters can be generated as a power combination of three parameters randomly or intentionally selected from parameter 316 and / or the difference parameters. Each of those three parameters selected from parameter 316 and / or the difference parameters can be raised to an exponent and can be multiplied or added to two other parameters that are similarly raised to an exponent. The powers raised to the same exponent for each of the three parameters selected from parameter 316 and / or the difference parameters can be, but do not need to be, the same.

[0045] In some implementations, for example, the combined parameters can be generated using a predetermined and finite range of exponents. For instance, in one such implementation, the exponent used to generate the combined parameters can be an integer power selected from -2, -1, zero, 1, and two (-2, -1, 0, 1, 2). Therefore, each combined parameter can take the following form:

[0046] X = Y1 a *Y2 b *...Y n c (Equation 1)

[0047] For example, each Y is one of the parameters in 316 or one of the difference parameters, n is any integer greater than 2, and each of a, b, and c can be any one of -2, -1, 0, 1, and 2. In one implementation, Equation 1 can be applied to the parameters 316, the difference parameters, and virtually all possible power combinations of the parameters 316 and the difference parameters, subject to the predetermined constraints discussed above, such as the value of n and the range of values ​​for the exponents that can be selected.

[0048] Flowchart 470 continues to analyze parameter 316, the difference parameter, and the combined parameter to identify a reduced set of parameters related to health status (e.g., related to hypotension in this exemplary method) (action 476). The analysis of parameter 316, the difference parameter, and the combined parameter used to identify the reduced set of parameters related to hypotension can be performed by the predictive risk model training software code 110 / 210a / 210b / 310, which is executed by the hardware processor 104 / 204 / 234 / 334.

[0049] As mentioned above, there can be thousands of parameters 316 and the difference parameters together. Therefore, and given the process of generating combined parameters described above, the combinations of parameters 316, the difference parameters, and the combined parameters can accumulate to millions. To make the analysis of such a large number of variables tractable, for example, in one embodiment, the analysis of parameters 316, the difference parameters, and the combined parameters can be performed as a receiver operating characteristic (ROC) analysis of these parameters.

[0050] ROC analysis is a method for illustrating the performance of binary classifiers because its discrimination threshold varies. ROC curves are created by plotting the true positive rate against the false positive rate at various threshold settings. The area under the ROC curve (AUC) can be used to judge the performance of different classifiers, and a higher AUC indicates a better classifier. On a predefined dataset of positive and negative results in healthy states, ROC analysis can be performed on each parameter to obtain its AUC. Parameters with large AUC values ​​are then retained.

[0051] The result of this ROC analysis is a reduced set of parameters associated with hypotension. For example, with a cumulative number of parameters 316, differential parameters, and combined parameters between two and three million, the reduced set could include fewer than about 200 parameters identified as associated with hypotension.

[0052] like Figure 4 As shown, flowchart 470 continues to identify parameter prediction set 114 / 214 from the parameter reduction set, which enables the prediction of the health status of a living subject, such as hypotension (action 477). The identification of parameter prediction set 114 / 214 from the parameter reduction set can be performed by the prediction risk model training software code 110 / 210a / 210b / 310 executed by the hardware processor 104 / 204 / 234 / 334.

[0053] In one implementation, the parameter prediction set 114 / 214 includes a subset of the parameter reduction set that has the strongest correlation with hypotension. For example, the correlation of each parameter included in the parameter reduction set can be determined by sequentially testing predictions of health states (e.g., hypotension) generated using each parameter in the parameter reduction set. In such an implementation, only those parameters from the parameter reduction set that have a measured correlation with hypotension that meets a threshold or cutoff correlation value are included as one of the parameters in the parameter prediction set 114 / 214.

[0054] In another implementation, machine learning techniques can be used to identify the parameter prediction set 114 / 214, such as sequential feature selection, whether forward or backward selection. Sequential feature selection involves adding or removing parameters one by one into or from the reduced parameter set in a machine learning model: a classification model or a regression model. Sequential feature selection aims to minimize the root mean square error (for regression models) or misclassification rate (for classification models) on all possible combinations of the reduced parameter set by adding (for forward selection) or removing (for backward selection) parameters one by one into or from the regression or classification model. Classification and regression models can include: linear regression, logistic regression, discriminant analysis, neural networks, support vector machines, nearest neighbors, classification and regression trees or ensemble methods such as random forests, to name just a few.

[0055] In another implementation, machine learning techniques such as optimal subset selection (leaps algorithm and boundary algorithm), ridge regression, lasso regression, minimum angle regression or principal component regression and partial least squares can be used to identify the parameter prediction set 114 / 214.

[0056] As a concrete example, where the parameter reduction set derived from the analysis of parameter 316, the difference parameters, and the combined parameters includes up to about 200 parameters, the parameter prediction set 114 / 214 can be of a number ranging from a few parameters (e.g., five or fewer) to up to about fifty parameters. Appendix A of this disclosure provides an exemplary, but not exhaustive, table listing exemplary cases for hypotension prediction using parameter prediction set 114 / 214 and exemplary sampling criteria associated with its determination. Predictive risk model training software code 110 / 210a / 210b / 310, executed by hardware processor 104 / 204 / 234 / 334, can utilize parameter prediction set 114 / 214 to significantly optimize predictive risk model 112 / 212 for predicting hypotension in living subjects.

[0057] Flowchart 470 may end with training the predictive risk models 112 / 212 to include some or all of the parameter prediction sets 114 / 214 using the aforementioned training subset (action 478). The predictive risk models 112 / 212 may include machine learning models: regression models or classification models. Examples of regression and classification models suitable for training the predictive risk models 112 / 212 may include linear regression, logistic regression, discriminant analysis, neural networks, support vector machines, nearest neighbor, classification and regression trees, or ensemble methods such as random forests, to name just a few.

[0058] The predictive risk model 112 / 212 may include all or a subset of the parameter prediction set 114 / 214, as well as additional parameters. The predictive risk model 112 / 212 may include model coefficients that must be determined through model training. Training the predictive risk model means using a numerical computation program to compute the predictive risk model coefficients to minimize the cost function representing the error of the predictive risk model output to the true value of the training subset.

[0059] As an example, a predictive risk model trained from action 478 using logistic regression can be represented as:

[0060] Risk Score = 1 / (1+e- A (Equation 2)

[0061] in:

[0062]

[0063] And among them:

[0064] ν1 = CWI, which is cardiac work indexed by the patient's body surface area;

[0065] ν2 = MAPavg, which is the averaged mean arterial pressure;

[0066] ν3 = ΔMAPavg, which is the averaged change in mean arterial pressure compared to the initial value;

[0067] ν4 = avgSysDec, which is the averaged pressure at the lower part of the contraction period;

[0068] ν5 = ΔSys, which is the change in systolic pressure compared to the initial value;

[0069] ν6 = ppAreaNor, which is the normalized area under the arterial pressure waveform;

[0070] ν7 = biasDia, which is the deviation of the diastolic slope;

[0071] ν8 = CW, which is the work done by the heart;

[0072] ν9 = mapDnlocArea, which is the area between the first instance of MAP and the dicrotic notch under the arterial pressure waveform;

[0073] ν 10 =SWcomb, which stands for SWcomb;

[0074] ν 11 =ppArea, which is the area under the arterial pressure waveform;

[0075] ν 12 =decAreaNor, which is the normalized area during the descent period;

[0076] ν 13 =slopeSys, which is the slope during the contraction phase;

[0077] ν 14 =Cwk, which stands for Windkessel (elastic cavity) compliance;

[0078] ν 15 =sys_rise_area_nor, which is the normalized area during the contraction rise phase;

[0079] ν 16 =pulsepres, i.e., pulse pressure;

[0080] ν 17 =avg_sys, which is the averaged pressure during the systolic phase;

[0081] ν 18 =dpdt2, which is the maximum value of the second derivative of the pressure waveform;

[0082] ν 19 =dpdt, which is the maximum value of the first derivative of the pressure waveform; and

[0083] Δ = the change in value over time compared to its initial value

[0084] c0, c1, ..., c 11 It is a constant coefficient.

[0085] Although not included in flowchart 470, in some embodiments, this method may also include updating the parameter prediction set 114 / 214 and the prediction risk model coefficients based on newly received vital sign data 460. That is, the hardware processor 104 / 204 / 234 / 334 may be configured to execute prediction risk model training software code 110 / 210a / 210b / 310 to update the parameter prediction set 114 / 214 after training the prediction risk model 112 / 212.

[0086] Therefore, this application discloses systems and methods for training predictive risk models for various potentially dangerous physiological states that critically ill or surgical patients may be susceptible to. Examples of these physiological states, as described above, include hypotension, hypovolemia, acute blood loss, septic shock, extubation failure, postoperative complications, and cardiovascular failure or collapse, to name just a few. According to various embodiments, the systems and methods disclosed in this application can be used by healthcare professionals to predict dangerous physical states before they manifest in living subjects. Due to this early warning, the systems and methods disclosed in this application enable the preparation of effective medical interventions for early treatment of or complete prevention of anticipated conditions.

[0087] As can be clearly seen from the above description, various techniques can be used to implement the concepts described in this application without departing from the scope of these concepts. Furthermore, although the concepts have been described with specific reference to certain embodiments, those skilled in the art will recognize that changes in form and detail can be made without departing from the scope of these concepts. Thus, the described embodiments are to be considered illustrative rather than restrictive in all respects. It should also be understood that this application is not limited to the specific embodiments described herein; rather, many rearrangements, modifications, and substitutions are possible without departing from the scope of this disclosure.

[0088] Appendix A

[0089]

[0090]

Claims

1. A system for training a risk prediction model, the system comprising: Hardware processor; System memory; and The predictive risk model training software code stored in the system memory; The hardware processor is configured to execute the predictive risk model training software code to: Arterial pressure data for low blood pressure were received from each subject in a subject population that included both positive and negative subjects. Define the dataset used to train the predicted risk model; The arterial pressure data is converted into a first plurality of parameters characterizing the arterial pressure data; The second plurality of difference parameters are obtained from the first plurality of parameters by determining at least one of the following: The change of the parameters in the first plurality of parameters relative to time; The change of the parameters in the first plurality of parameters relative to the frequency; and The change of a parameter among the first plurality of parameters relative to the other parameters among the first plurality of parameters; A third set of combined parameters is generated using the first plurality of parameters and the second plurality of difference parameters. Each of the third plurality of combined parameters is calculated using a total of three parameters, each of which is selected from one or more of the first plurality of parameters and the second plurality of difference parameters; and Each of the third plurality of combined parameters comprises a power combination of a subset of the first plurality of parameters and the second plurality of difference parameters, the power combination comprising integer powers from negative two, negative one, one, and two (-2, -1, 1, 2); The third set of combined parameters is analyzed to identify a reduced set of parameters associated with the low blood pressure. Identifying a parameter prediction set from the reduced set of said parameters to predict hypotension in a living subject, wherein identifying said parameter prediction set includes: Use a regression model to identify the parameter prediction set as a subset of the reduced set of parameters that minimizes the root mean square error of the regression model; or The classification model is used to identify the parameter prediction set as a subset of the reduced set of parameters that minimizes the misclassification rate of the classification model; and The prediction risk model is trained by calculating prediction risk model coefficients based on the parameter prediction set, wherein calculating the prediction risk model coefficients includes minimizing a cost function representing the error of the prediction risk model's output to the true value of the dataset used to train the prediction risk model.

2. The system according to claim 1, wherein the parameter in the first plurality of parameters represents: Characteristics of arterial pressure waveforms; The behavior of the arterial pressure waveform during the interval defined by the characteristics of the arterial pressure waveform; or The characteristics of the arterial pressure waveform and the behavior of the arterial pressure waveform during the interval.

3. The system according to claim 2, wherein the differential parameters in the second plurality of differential parameters are different from the parameters in the first plurality of parameters.

4. The system of claim 1, wherein the combined parameters in the third plurality of combined parameters represent a hemodynamic complexity measurement, the hemodynamic complexity measurement quantifying the regularity and unpredictability of fluctuations in cardiac measurements.

5. The system of claim 1, wherein a reduced set of the parameters associated with the hypotension is identified by receiver operating characteristic analysis, i.e., ROC analysis, of the third plurality of combined parameters.

6. The system of claim 1, wherein using the regression model to identify the parameter prediction set comprises: On a parameter-by-parameter basis, each parameter from the reduced set of the parameters is added to the regression model to identify the subset of the reduced set of the parameters that minimizes the root mean square error of the regression model; or On a parameter-by-parameter basis, each parameter from the reduced set of the parameters is removed from the regression model to identify the subset of the reduced set of the parameters that minimizes the root mean square error of the regression model.

7. The system of claim 1, wherein using the classification model to identify the parameter prediction set comprises: On a parameter-by-parameter basis, each parameter from the reduced set of the parameters is added to the classification model to identify the subset of the reduced set of the parameters that minimizes the misclassification rate of the classification model; or On a parameter-by-parameter basis, each parameter from the reduced set of the parameters is removed from the classification model to identify the subset of the reduced set of the parameters that minimizes the misclassification rate of the classification model.

8. A method for training a predictive risk model using a system, said system comprising a hardware processor and predictive risk model training software code stored in system memory, said method comprising: The software code for training the predictive risk model, executed by the hardware processor, receives arterial pressure data for low blood pressure from each subject in a subject population that includes both positive and negative subjects. The dataset used to train the predictive risk model is defined by the software code for training the predictive risk model, which is executed by the hardware processor. The arterial pressure data is converted into a first plurality of parameters characterizing the arterial pressure data by training software code for the predictive risk model executed by the hardware processor. The software code for training the predictive risk model, executed by the hardware processor, obtains a second plurality of differencing parameters from the first plurality of parameters by determining at least one of the following: The change of the parameters in the first plurality of parameters relative to time; The change of the parameter in the first plurality of parameters relative to frequency; and The change of the parameter in the first plurality of parameters relative to the other parameters in the first plurality of parameters; The software code for training the predictive risk model, executed by the hardware processor, generates a third set of combined parameters using the first plurality of parameters and the second plurality of differential parameters. Each of the third plurality of combined parameters is calculated using a total of three parameters, each of which is selected from one or more of the first plurality of parameters and the second plurality of difference parameters; and Each of the third plurality of combined parameters comprises a power combination of a subset of the first plurality of parameters and the second plurality of difference parameters, the power combination comprising integer powers from negative two, negative one, one, and two (-2, -1, 1, 2); The software code for training the predictive risk model, executed by the hardware processor, is analyzed to identify a reduced set of parameters associated with the low blood pressure. The software code, trained by the predicted risk model and executed by the hardware processor, identifies a set of parameter predictions from the reduced set of parameters to predict hypotension in a living subject, wherein identifying the set of parameter predictions includes: Use a regression model to identify the parameter prediction set as a subset of the reduced set of parameters that minimizes the root mean square error of the regression model; or The classification model is used to identify the parameter prediction set as a subset of the reduced set of parameters that minimizes the misclassification rate of the classification model; and The predictive risk model is trained by calculating predictive risk model coefficients based on the parameter prediction set using software code executed by the hardware processor, wherein calculating the predictive risk model coefficients includes minimizing a cost function representing the error of the output of the predictive risk model to the true value of the dataset used to train the predictive risk model.

9. The method of claim 8, wherein the parameter in the first plurality of parameters represents: Characteristics of arterial pressure waveforms; The behavior of the arterial pressure waveform during the interval defined by the characteristics of the arterial pressure waveform; or The characteristics of the arterial pressure waveform and the behavior of the arterial pressure waveform during the interval.

10. The method of claim 9, wherein the differential parameters in the second plurality of differential parameters are different from the parameters in the first plurality of parameters.

11. The method of claim 8, wherein the combined parameters in the third plurality of combined parameters represent a hemodynamic complexity measurement, the hemodynamic complexity measurement quantifying the regularity and unpredictability of fluctuations in cardiac measurements.

12. The method of claim 8, wherein analyzing the third plurality of combined parameters to identify a reduced set of the parameters associated with the hypotension includes performing a receiver operating characteristic (ROC) analysis of the third plurality of combined parameters.

13. The method of claim 8, wherein using the regression model to identify the parameter prediction set comprises: On a parameter-by-parameter basis, each parameter from the reduced set of the parameters is added to the regression model to identify the subset of the reduced set of the parameters that minimizes the root mean square error of the regression model; or On a parameter-by-parameter basis, each parameter from the reduced set of the parameters is removed from the regression model to identify the subset of the reduced set of the parameters that minimizes the root mean square error of the regression model.

14. The method of claim 8, wherein using the classification model to identify the parameter prediction set comprises: On a parameter-by-parameter basis, each parameter from the reduced set of the parameters is added to the classification model to identify the subset of the reduced set of the parameters that minimizes the misclassification rate of the classification model; or On a parameter-by-parameter basis, each parameter from the reduced set of the parameters is removed from the classification model to identify the subset of the reduced set of the parameters that minimizes the misclassification rate of the classification model.

15. A computer-readable non-transitory medium having instructions stored thereon, wherein, when executed by a hardware processor, the instructions are instantiated using a method comprising the steps of: Arterial pressure data for low blood pressure were received from each subject in a subject population that included both positive and negative subjects. Define the dataset used to train the risk prediction model; The arterial pressure data is converted into a first plurality of parameters characterizing the arterial pressure data; The second plurality of difference parameters are obtained from the first plurality of parameters by determining at least one of the following: The change of the parameters in the first plurality of parameters relative to time; The change of the parameter in the first plurality of parameters relative to frequency; and The change of the parameter in the first plurality of parameters relative to the other parameters in the first plurality of parameters; A third set of combined parameters is generated using the first plurality of parameters and the second plurality of difference parameters. Each of the third plurality of combined parameters is calculated using a total of three parameters, each of which is selected from one or more of the first plurality of parameters and the second plurality of difference parameters; and Each of the third plurality of combined parameters comprises a power combination of a subset of the first plurality of parameters and the second plurality of difference parameters, the power combination comprising integer powers from negative two, negative one, one, and two (-2, -1, 1, 2); The third set of combined parameters is analyzed to identify a reduced set of parameters associated with the low blood pressure. Identifying a parameter prediction set from the reduced set of parameters to predict the hypotension in a living subject, wherein identifying the parameter prediction set includes: Use a regression model to identify the parameter prediction set as a subset of the reduced set of parameters that minimizes the root mean square error of the regression model; or The classification model is used to identify the parameter prediction set as a subset of the reduced set of parameters that minimizes the misclassification rate of the classification model; and The prediction risk model is trained by calculating prediction risk model coefficients based on the parameter prediction set, wherein calculating the prediction risk model coefficients includes minimizing a cost function representing the error of the prediction risk model's output to the true value of the dataset used to train the prediction risk model.

16. The computer-readable non-transitory medium of claim 15, wherein the parameter in the first plurality of parameters represents: Characteristics of arterial pressure waveforms; The behavior of the arterial pressure waveform during the interval defined by the characteristics of the arterial pressure waveform; or The characteristics of the arterial pressure waveform and the behavior of the arterial pressure waveform during the interval.

17. The computer-readable non-transitory medium of claim 16, wherein the differential parameters in the second plurality of differential parameters are different from the parameters in the first plurality of parameters.

18. The computer-readable non-transitory medium of claim 15, wherein the combined parameters in the third plurality of combined parameters represent a hemodynamic complexity measurement that quantifies the regularity and unpredictability of fluctuations in cardiac measurements.

19. The computer-readable non-transitory medium of claim 18, wherein analyzing the third plurality of combined parameters to identify a reduced set of the parameters associated with the hypotension includes performing a receiver operating characteristic analysis, i.e., a ROC analysis, of the third plurality of combined parameters.

20. The computer-readable non-transitory medium of claim 15, wherein using the regression model to identify the parameter prediction set comprises: On a parameter-by-parameter basis, each parameter from the reduced set of the parameters is added to the regression model to identify the subset of the reduced set of the parameters that minimizes the root mean square error of the regression model; or On a parameter-by-parameter basis, each parameter from the reduced set of the parameters is removed from the regression model to identify the subset of the reduced set of the parameters that minimizes the root mean square error of the regression model.

21. The computer-readable non-transitory medium of claim 15, wherein using the classification model to identify the parameter prediction set comprises: On a parameter-by-parameter basis, each parameter from the reduced set of the parameters is added to the classification model to identify the subset of the reduced set of the parameters that minimizes the misclassification rate of the classification model; or On a parameter-by-parameter basis, each parameter from the reduced set of the parameters is removed from the classification model to identify the subset of the reduced set of the parameters that minimizes the misclassification rate of the classification model.

22. The system of claim 1, wherein the first plurality of parameters includes one or more of cardiac output, cardiac index, stroke volume, stroke volume index, pulse rate, systemic vascular resistance, systemic vascular resistance index, baroreflex sensitivity measurement, hemodynamic complexity measurement and / or frequency domain hemodynamic characteristics.

23. The method of claim 8, wherein the first plurality of parameters includes one or more of cardiac output, cardiac index, stroke volume, stroke volume index, pulse rate, systemic vascular resistance, systemic vascular resistance index, baroreflex sensitivity measurement, hemodynamic complexity measurement and / or frequency domain hemodynamic characteristics.

24. The computer-readable non-transitory medium of claim 15, wherein the first plurality of parameters includes one or more of cardiac output, cardiac index, stroke volume, stroke volume index, pulse rate, systemic vascular resistance, systemic vascular resistance index, baroreflex sensitivity measurement, hemodynamic complexity measurement and / or frequency domain hemodynamic characteristics.

25. A method for training a risk prediction model, comprising: The software code is trained by a predictive risk model executed by a hardware processor, and receives arterial pressure data on low blood pressure for each subject in a subject population that includes both positive and negative subjects. The dataset used to train the predictive risk model is defined by the software code for training the predictive risk model, which is executed by the hardware processor. The arterial pressure data is converted into a first plurality of parameters characterizing the arterial pressure data by training software code for the predictive risk model executed by the hardware processor. The software code for training the predictive risk model, executed by the hardware processor, obtains a second plurality of differencing parameters from the first plurality of parameters by determining at least one of the following: The change of the parameters in the first plurality of parameters relative to time; The change of the parameter in the first plurality of parameters relative to frequency; and The change of the parameter in the first plurality of parameters relative to the other parameters in the first plurality of parameters; The software code for training the predictive risk model, executed by the hardware processor, generates a third set of combined parameters using the first plurality of parameters and the second plurality of differential parameters. Each of the third plurality of combined parameters is calculated using a total of three parameters, each of which is selected from one or more of the first plurality of parameters and the second plurality of difference parameters; and Each of the third plurality of combined parameters comprises a power combination of a subset of the first plurality of parameters and the second plurality of difference parameters, the power combination comprising integer powers from negative two, negative one, one, and two (-2, -1, 1, 2); The software code for training the predictive risk model, executed by the hardware processor, is analyzed to identify a reduced set of parameters associated with hypotension in a living subject. The software code for training the predicted risk model, executed by the hardware processor, identifies a set of parameter predictions from the reduced set of parameters to predict hypotension in the living subject, wherein identifying the set of parameter predictions includes: Use a regression model to identify the parameter prediction set as a subset of the reduced set of parameters that minimizes the root mean square error of the regression model; or The classification model is used to identify the parameter prediction set as a reduced set of parameters that minimizes the misclassification rate of the classification model; The predictive risk model is trained by the software code executing the predictive risk model, which calculates predictive risk model coefficients based on the parameter prediction set to generate a trained predictive risk model. Calculating the predictive risk model coefficients includes minimizing a cost function representing the error of the predictive risk model's output to the true values ​​of the dataset used to train the predictive risk model. The trained predictive risk model was used to predict low blood pressure in the living subject.

Images