Machine-learning-based prediction of adverse outcomes using data from implantable or wearable cardiac devices

Abstract

A method for machine-learning-based prediction of adverse outcomes using measured time-varying values of physiological parameters generated from output of an implantable or wearable cardiac device includes receiving, as input to a trained machine learning model, measured time-varying values of physiological parameters generated from output of one or more embedded sensors of a wearable or implantable cardiac device worn by or implanted within an individual subject. The method further includes generating, as output from the trained machine learning model, a value indicating a personalized risk estimate of all-cause mortality or a composite event of all-cause mortality or heart failure hospitalization for the individual subject. The method further includes performing, based on the value indicating the personalized risk estimate of all-cause mortality or the composite event, an intervention for the individual subject.

Description

[0001] METHODS, SYSTEMS, AND COMPUTER READABLE MEDIA FOR MACHINE-LEARNING-BASED PREDICTION OF ADVERSE OUTCOMES USING DATA FROM IMPLANTABLE OR WEARABLE CARDIAC DEVICES

[0002] GOVERNMENT INTEREST

[0003] This invention was made with government support under Grant Number HL141644 awarded by the National Institutes of Health. The government has certain rights in the invention.

[0004] PRIORITY CLAIM

[0005] This application claims the priority benefit of U.S. Provisional Patent Application Serial No. 63 / 684,842 filed August 19, 2024, the disclosure of which is incorporated herein by reference in its entirety.

[0006] TECHNICAL FIELD

[0007] The subject matter described herein relates to predicting adverse outcomes, such as all-cause mortality and heart failure hospitalization. More particularly, the subject matter described herein relates to methods, systems, and computer readable media for machine-learning-based prediction of allcause mortality and a composite event of all-cause mortality and heart failure hospitalization using time-varying data from implantable or wearable cardiac devices.

[0008] BACKGROUND

[0009] Predicting the clinical trajectory of individual patients with implantable cardioverter defibrillators (ICDs) is essential to inform clinical care. Machine learning approaches can potentially overcome the limitations of conventional statistical methods and provide more accurate, personalized, risk estimates. Existing risk stratification approaches were developed with a limited set of predictor variables and used conventional statistical approaches (e.g., regression models) that assume linear relationships between predictors, have difficulty managing complex interactions, and have limited tolerance for missing data, which is common in clinical practice. Existing models were also developed in highly selected patients (e.g., ‘healthier’ predominately white patients in clinical trials) and have not been externally validated in real-world populations, which is critical to determine the generalizability of the risk model to new patients and different clinical settings. Some models also fail to predict all-cause mortality and / or heart failure hospitalization and instead predict sudden cardiac arrest. Another shortcoming of existing risk scores is that they do not quantify the relative importance of risk factors that lead to the predicted outcome. Yet another shortcoming of existing models is the inability to predict a risk that is associated with a particular time period for an individual subject.

[0010] Accordingly, in light of these and other difficulties, there exists a need for improved methods, systems, and computer readable media for predicting adverse outcomes, such as all-cause mortality and heart failure hospitalization, from time varying data recorded by an implantable or wearable cardiac measurement device.

[0011] SUMMARY

[0012] A method for machine-learning-based prediction of adverse outcomes using measured time-varying values of physiological parameters generated from output of one or more embedded sensors of an implantable or wearable cardiac device worn by or implanted within an individual subject includes, receiving as input to a trained machine learning model, the measured timevarying values of the physiological parameters generated from output of the one or more embedded sensors of the implantable or wearable cardiac device. The method further includes generating, as output from the trained machine learning model, a value indicating a personalized risk estimate of allcause mortality or a composite event of all-cause mortality or heart failure hospitalization for the individual subject. The method further includes performing, based on the value indicating the personalized risk estimate of all- cause mortality or the composite event, an intervention for the individual subject.

[0013] The term “wearable cardiac device” includes medical grade devices with regulatory approval for heart rate monitoring and / or regulation as well as wearable consumer grade devices with the capability to monitor heart rate.

[0014] According to another aspect of the subject matter described herein, the wearable or implantable cardiac device comprises a cardiovascular implantable electronic device (CIED).

[0015] According to another aspect of the subject matter described herein, the CIED comprises an implantable cardioverter defibrillator (ICD) with or without cardiac resynchronization therapy (CRT-D), an insertable cardiac monitor (ICM), a pulmonary pressure-sensing device, a cardiac loop recorder, or a pacemaker with or without cardiac resynchronization therapy (CRT-P).

[0016] According to another aspect of the subject matter described herein, the wearable or implantable cardiac device comprises a wearable device.

[0017] According to another aspect of the subject matter described herein, the wearable device comprises a wearable heart rate monitor.

[0018] According to another aspect of the subject matter described herein, the wearable device comprises a smart watch or a smart ring.

[0019] According to another aspect of the subject matter described herein, the method for machine learning prediction of adverse outcomes using timevarying values of physiological parameters generated from output of one or more embedded sensors of an implantable or wearable cardiac device includes outputting, by the trained machine learning model, parameter importance values indicating importances of the parameters contributing to the value indicating the personalized risk estimate of all-cause mortality or the composite event for the individual subject.

[0020] According to another aspect of the subject matter described herein, the value indicating the personalized risk estimate of all-cause mortality or the composite event of all-cause mortality or heart failure hospitalization for the individual subject is associated with a time period when the all-cause mortality or a composite event of all-cause mortality or heart failure hospitalization is likely to occur.

[0021] According to another aspect of the subject matter described herein, the trained machine learning model comprises a random forest for survival, longitudinal and multivariate (RF-SLAM) model.

[0022] According to another aspect of the subject matter described herein, the RF-SLAM model is trained using measured values of time-varying wearable or implantable cardiac device parameters for subjects in a study cohort.

[0023] According to another aspect of the subject matter described herein, performing the intervention includes performing replacement, repair, or maintenance of the wearable or implantable cardiac device.

[0024] According to another aspect of the subject matter described herein, performing the intervention includes generating a cardiac treatment plan tailored to reduce the value indicating the risk of all-cause mortality or the composite event.

[0025] According to another aspect of the subject matter described herein, a system for machine-learning-based prediction of adverse outcomes using measured time-varying values of physiological parameters generated from output of one or more embedded sensors of an implantable or wearable cardiac device is provided. The system includes a computing platform including at least one processor and a memory. The system further includes a trained machine learning model comprising computer-executable instructions stored in the memory and executable by the at least one processor for receiving, as input, measured time-varying values of physiological parameters generated from output of one or more embedded sensors of a wearable or implantable cardiac device worn by or implanted within an individual subject and generating, as output, a value indicating a personalized risk estimate of all-cause mortality or a composite event of allcause mortality or heart failure hospitalization for the individual subject.

[0026] According to another aspect of the subject matter described herein, the trained machine learning model is configured to generate, as output, parameter importance values indicating importances of the parameters contributing to the value indicating the personalized risk estimate of all-cause mortality or the composite event for the individual subject.

[0027] According to another aspect of the subject matter described herein, a non-transitory computer readable medium having stored thereon executable instructions that when executed by a processor of a computer controls the computer to perform steps is provided. The steps include receiving, as input to a trained machine learning model, measured time-varying values of physiological parameters generated from output of one or more embedded sensors of a wearable or implantable cardiac device worn by or implanted within an individual subject. The steps further include generating, as output from the trained machine learning model, a value indicating a personalized risk estimate of all-cause mortality or a composite event of all-cause mortality or heart failure hospitalization for the individual subject.

[0028] The subject matter described herein can be implemented in software in combination with hardware and / or firmware. For example, the subject matter described herein can be implemented in software executed by a processor. In one exemplary implementation, the subject matter described herein can be implemented using a non-transitory computer readable medium having stored thereon computer executable instructions that when executed by the processor of a computer control the computer to perform steps. Exemplary computer readable media suitable for implementing the subject matter described herein include non-transitory computer-readable media, such as disk memory devices, chip memory devices, programmable logic devices, and application specific integrated circuits. In addition, a computer readable medium that implements the subject matter described herein may be located on a single device or computing platform or may be distributed across multiple devices or computing platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Exemplary implementations of the subject matter described herein will now be explained with reference to the accompanying drawings, of which:

[0030] Figure 1 is a flow chart of study participants;

[0031] Figures 2A-2D are graphs of relative importance values of the variables in the machine learning models in predicting the indicated outcomes. Figures 2A-2D illustrate the most important predictors for identified by the RF-SLAM model for predicting (Figure 2A) 3-month death risk; (Figure 2B) 1 -year death risk; (Figure 2C) 3-month risk of death or heart failure (HF) hospitalization; (Figure 2D) 1 -year risk of death or HF hospitalization. The bars represent the relative contribution of individual variables (percentage of trees used by the model) in predicting the outcome. All models are consistent in identifying timevarying device measurements as the most important predictors of death and death / HF hospitalization;

[0032] Figure 3 illustrates time-varying area under the ROC curve for the RF- SLAM Models compared to the traditional Cox Model for 3-month risk of (a) death or (b) death or HF hospitalization. Predictors used in > 15% of the trees in the RF-SLAM model 2 were included in the time-varying Cox model. They were (a) physical activity, impedance, HRV, age, number of shocks, HF hospitalization, spontaneous VT / VF episodes, number of ATP sequences, night heart rate, time in AT / AF, atrial pace %, ventricular pace %, spontaneous NST episodes, baseline CKD, daytime heart rate, BMI and smoking status for 3-month risk of death; (b) impedance, physical activity, HRV, time in AT / AF, night hear rate, baseline heart failure, baseline diuretics, baseline CKD, spontaneous VT / VF episodes, number of shocks, number of ATP sequences, atrial pace %, daytime heart rate, ventricular pace %, baseline LVEF, age and baseline atrial flutter / fibrillation for 3-month risk of death or HF hospitalization;

[0033] Figure 4 illustrates dependence plots for top predictors of 3-month risk of death. In Figure 4, Individual data points versus the predicted risk are shown, and the red line shows the relationship between 3-month risk of death and the predictor variables. The order of each predictor is consistent with the variable importance ranking in Figure 2A, beginning with the predictor (physical activity) most frequently selected by trees;

[0034] Figure 5 illustrates dependence plots for top predictors of 1 year risk of death. In Figure 5, individual data points versus the predicted risk are shown, and the red line shows the relationship between 1 -year risk of death and the predictor variables. The order of each predictor is consistent with the variable importance ranking in Figure 2B, beginning with the predictor (physical activity) most frequently selected by trees;

[0035] Figure 6 illustrates dependence plots for top predictors of 3-month risk of death or HF hospitalization. In Figure 6, individual data points versus the predicted risk are shown, and the red line shows the relationship between 3- month risk of death / HF hospitalization and the predictor variables. The order of each predictor is consistent with the variable importance ranking in Figure 2C, beginning with the predictor (thoracic impedance) most frequently selected by trees;

[0036] Figure 7 illustrates dependence plots for top predictors of 1 year risk of death or HF hospitalization. In Figure 7, individual data points versus the predicted risk are shown, and the red line shows the relationship between 1- year risk of death / HF hospitalization and the predictor variables. The order of each predictor is consistent with the variable importance ranking in Figure 2D, beginning with the predictor (thoracic impedance) most frequently selected by trees;

[0037] Figure 8 illustrates calibration of RF-SLAM Model 2 for 3-month risk of death in the VHA training cohort. In Figure 8, the calibration plot shows the observed probability versus predicted probability for the risk of death in 3 months. For the highest risk group, a minimal overestimation of predicted risks was observed. The histogram shows a dispersed bimodal distribution, with 81 .6% of person-time intervals having a predicted probability of death less than 0.01 and 10.4% of person-time intervals having a predicted probability between 0.03-0.0375. Despite this overestimation, the calibration intercept (0.00; 95%CI -0.06, 0.06) and slope (0.99; 95%CI 0.94, 1.04) still indicated very good overall calibration;

[0038] Figure 9 illustrates calibration of RF-SLAM Model 2 for 1 -year risk of death in the VHA training cohort. In Figure 9, the calibration plot shows the observed probability versus predicted probability for the risk of death in 1 -year. The predicted risk ranges from 0.001 to 0.165. The histogram at the bottom also shows a dispersed bimodal distribution, with 64.2% of person-time intervals having a predicted probability of death less than 0.025 and 12.8% of person-time intervals having predicted probability between 0.08-0.10. The model tends to overestimate the risk for observations with higher predicted risk, especially beyond predicted risk of 0.10. However, only 1.6% of the observations have a predicted risk greater than 0.10;

[0039] Figure 10 illustrates calibration of RF-SLAM Model 2 for 3-month risk of death or heart failure hospitalization in the VHA training cohort. The predicted risk of death or heart failure hospitalization in 3 months ranges from 0.0007 to 0.169. The histogram at the bottom shows a bimodal distribution of predicted probability, with 71.3% of person-time intervals having a predicted probability less than 0.015 and 13.0% of person-time intervals having a predicted probability between 0.03-0.05. The model tends to underestimate the risk for observations predicted to have higher risk of heart failure hospitalization or death in 3 months. This underestimation is likely due to the small sample size of high-risk observations, as only 0.15% of the observations have a predicted risk greater than 0.1 ;

[0040] Figure 11 illustrates calibration of RF-SLAM Model 2 for 1 -year risk of death or heart failure hospitalization in the VHA training cohort. The predicted risk of death or heart failure hospitalization in 1-year ranges from 0.007 to 0.342, showing a bimodal distribution of predicted probabilities, which indicates that the model can effectively discriminate between high- and low- risk observations. The calibration curve aligns with the diagonal line, indicating very good calibration for all risk groups; Figure 12 illustrates calibration of the RF-SLAM model 2 for (a) 3-month risk of death, (b) 1 -year risk of death, (c) 3-month risk of death or HF hospitalization and (d) 1-year risk of death or HF hospitalization in the UNC validation cohort. The plots indicate excellent model calibration in the UNC validation cohort;

[0041] Figure 13 is a block diagram illustrating an exemplary system for machine-learning-based prediction of adverse outcomes using measured time-varying values of physiological parameters from one or more embedded sensors of an implantable or wearable cardiac device; and

[0042] Figure 14 is a flow chart illustrating an exemplary process for machine- learning-based prediction of adverse outcomes using measured time-varying values of physiological parameters generated from output of one or more embedded sensors of an implantable or wearable cardiac device.

[0043] DETAILED DESCRIPTION

[0044] The following is a description of a study in which a machine learning model is trained to take time varying data from an implantable cardioverter defibrillator well as patient demographic data and output an indicator of allcause mortality and a composite event of all-cause mortality and heart failure hospitalization.

[0045] Background: Predicting the clinical trajectory of individual patients with implantable cardioverter defibrillators (ICDs) is essential to inform clinical care. Machine learning approaches can potentially overcome the limitations of conventional statistical methods and provide more accurate, personalized, risk estimates.

[0046] Objectives: We sought to develop and externally validate a novel machine learning algorithm for predicting all-cause mortality and / or heart failure (HF) hospitalization in ICD patients with and without chronic resynchronization therapy (CRT) using variables that are readily available to treating clinicians. We also sought to identify key factors that separate patients along a continuum of risk.

[0047] Methods: Random forest for survival, longitudinal, and multivariate (RF- SLAM) data analysis was applied to predict 3-month and 1-year risks for allcause mortality and a composite outcome of death / HF hospitalization during the first 5 years of device implant. Models were trained using a nationwide cohort from the Veterans Health Administration. Three models were sequentially tested, and external validation was performed in a separate nonveteran clinical registry.

[0048] Results: The training and validation cohorts included 12,043 patients (age 67.5 ± 9.4 years) and 1 ,394 patients (age 66.3 ± 11.9 years), respectively. Median follow-up was 3.3 years for the training cohort and 3.6 years for validation cohort. The most accurate models for both outcomes included baseline demographics entered at the time of ICD implant (age, sex, CRT therapy) and time-varying ICD data with AUCs for predicting death at 3 months [0.91 , 95% confidence interval (Cl), 0.87 to 0.94] and 1 year [0.80 (Cl, 0.78 to 0.82)]; death / HF hospitalization at 3 months [0.81 (Cl, 0.79 to 0.83)] and 1 year [0.71 (Cl, 0.70 to 0.72)]. Models demonstrated high discrimination and good calibration in the validation cohort. Additionally, time-varying physiologic data from ICDs, especially daily physical activity, had substantial importance in predicting outcomes.

[0049] Conclusions: The RF-SLAM algorithm accurately predicted all-cause mortality and death / HF hospitalization at 3 months and 1 year during the first 5 years of device implant, demonstrating good internal and external validity. Prospective studies and randomized trials are needed to validate model performance in other populations and settings and to evaluate its impact on patient outcomes. Abbreviations

[0050] ICD = implantable cardioverter defibrillator

[0051] HF = heart failure

[0052] ML = machine learning

[0053] EHR = electronic health record

[0054] CRT-D = cardiac resynchronization therapy defibrillator

[0055] RF-SLAM = random forests for survival, longitudinal and multivariate data algorithm

[0056] VHA = Veterans health administration

[0057] UNC = University of North Carolina

[0058] Introduction

[0059] In the United States, more than 1.3 million patients are currently living with an implantable cardioverter defibrillator (ICD),1and this number is expected to rise as the population ages.2Patients with ICDs are a heterogeneous group who vary in their disease progression and risk for hospitalization and death.3A substantial proportion of ICD patients also have heart failure (HF) which carries a high burden of morbidity and mortality.4Thus, precise risk assessment is crucial for identifying high-risk individuals who are most likely to benefit from preventative interventions and facilitating clinician-patient risk communication and shared decision-making.

[0060] Several risk stratification tools have been developed to predict adverse events in patients with ICDs using clinical parameters from the time of implant and / or continuous physiologic sensor data automatically collected by the device (e.g., treated ventricular arrhythmias, heart rate, atrial fibrillation burden, physical activity).5-10However, their utility and adoption in clinical settings has been limited due to suboptimal performance, bias in the development cohorts, and poor generalizability. Most existing risk prediction tools were developed with a limited set of predictor variables and used conventional statistical approaches (e.g., regression models) that assume linear relationships between predictors, have difficulty managing complex interactions, and have limited tolerance for missing data, which is common in clinical practice. Existing models were also developed in highly selected patients (e.g., ‘healthier’ predominately white patients in clinical trials) and have not been externally validated in real-world populations, which is critical to determine the generalizability of the risk model to new patients and different clinical settings.

[0061] To address the limitations of previous models, we applied a contemporary machine learning (ML)-based approach [random forest for survival, longitudinal, and multivariate (RF-SLAM) data analysis]11to develop and externally validate a novel prediction algorithm to provide individualized or personalized risk estimates of all-cause mortality and hospitalization for HF using longitudinal data from two well-characterized, real-world registries of patients with ICDs (total N= 13,458). ML methods can improve risk prediction by analyzing large volumes of patient-level data and examining complex, nonlinear relationships between fixed characteristics and time-varying predictors, including interim clinical events.12ML models also have the computational power to examine individual risk trajectories across a long follow-up time and uncover hidden patterns and interactions between predictors, making them valuable tools for risk stratification in diverse populations. Therefore, the primary objective of this study was to build a pragmatic, potentially scalable, risk prediction tool with excellent discriminative ability using variables that are routinely collected in ICD patients. We also sought to identify key factors that separate patients along a continuum of risk.

[0062] Methods

[0063] Data Source and Study Cohorts

[0064] This study followed the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) reporting guideline.13Institutional Review Boards at each site approved the study protocol and waiver of informed consent. No funding or other research support was provided by the device manufacturers. All authors take responsibility for the integrity of the data and analyses. The data that support the findings of this study are available from the corresponding author upon reasonable request.

[0065] The training cohort included a nationwide sample from the Veterans Health Administration (VHA) National Cardiac Device Surveillance Program (NCDSP) (age > 18 years) who received a Medtronic implantable cardioverter defibrillator (ICD) or cardiac resynchronization therapy defibrillator (CRT-D) for primary or secondary prevention between January 1 , 2010 and August 16, 2020 and is capable of recording physical activity from a built-in accelerometer.14Daily physiologic data from ICDs were obtained directly from the device manufacturer and included the number of spontaneous ventricular tachycardia / ventricular fibrillation episodes, number of non-sustained ventricular tachycardia episodes, treated ventricular arrhythmias (antitachycardia pacing [ATP] and ICD shocks), daytime and nighttime heart rate, heart rate variability, minutes of atrial fibrillation, percentage of biventricular pacing, changes in intrathoracic impedance, and minutes of physical activity. Additional data from the time of implant (e.g., ejection fraction), demographics, and clinical information (inpatient, outpatient, and pharmacy data; date of death) were obtained from the VHA Corporate Data Warehouse.15Patients diagnosed with inherited channelopathies (e.g., Long QT Syndrome, Brugada syndrome), with a left ventricular assist device (LVAD), or with physical disabilities that interfere with performing mild to moderate-intensity physical activities (e.g., paralysis, lower extremity amputation, being wheelchair- bound), were excluded from the analysis as daily physical activity data was a key parameter in our models and has been associated with morbidity and mortality outcomes in prior studies of ICD patients.16

[0066] External validation was performed in a cohort of patients enrolled in the UNC Cardiovascular Device Surveillance Registry (UNC CDSR) who had received a Medtronic ICD or CRT-D between January 1 , 2010 and January 23, 2021. The UNC CDSR17 18is an ongoing prospective, clinical research registry of patients who have received pacemakers, ICDs, and CRT devices at the University of North Carolina (UNC) Medical Center and 10 affiliated hospitals located throughout central North Carolina. The registry collects daily device data from all remote monitoring transmissions and routine follow-up clinic visits for all consecutive implanted cardiac device implants, upgrades, and replacements. Device data are deterministically linked to patient-level information on sociodemographics, clinical history, and medications which are routinely abstracted from electronic health records (EHR) using standard procedures and validated ICD-9 and ICD-10 codes.19Patients who had an LVAD or were diagnosed with inherited channelopathies were excluded. Outcomes

[0067] The primary outcomes of interest were all-cause mortality and a composite outcome of all-cause mortality and heart failure (HF) hospitalization. These outcomes were selected to facilitate comparisons between our ML-based model and existing risk prediction tools.20 21For the training cohort, date of death within VA or outside hospitals was ascertained from the Social Security Administration and Center for Medicare and Medicaid Services file with accuracy comparable to National Death Index.15For the validation cohort, date of death was obtained from the EHR and from device clinic records. In both cohorts, a primary HF hospitalization was defined as any ICD-9-CM or ICD-10-CM HF discharge code used as the primary designated diagnostic code [ICD-9 codes: 402.X1 , 404. X1 , 404. X3, 428. X; or I, 2022, codes: 111.0, 113.0, 113.2, I50.X].22Patients were followed from baseline (defined as 60 days after device implantation to account for post- procedural recovery) until the earliest occurrence of an event, loss to followup (defined as the date of the last encounter), or end of follow-up (October 15th 2020 for training cohort and January 1st 2022 for validation cohort), whichever happened first.

[0068] Candidate Predictors

[0069] A wide range of candidate predictors were pre-specified based on existing literature,21 23clinical judgment of the authors, and availability in the cohorts. We included both time-fixed baseline characteristics and time-varying predictors which were binned into 3-month intervals. This timeframe was selected to coincide with the frequency of routine remote interrogations for ICDs / CRTs recommended by the current guidelines.24Demographics (age, sex, race, ethnicity, employment status, smoking status, body mass index), ejection fraction, CRT therapy, clinical comorbidities, and medications were obtained at the time of implant. Because social determinants of health (SDOH) have been shown to improve cardiovascular risk prediction in non-white individuals25and are associated with HF morbidity and related hospitalizations,26we included the Social Deprivation Index (SDI) in our models.27The SDI is a well-validated composite index of seven demographic and socioeconomic factors from the 2015 American Community Survey. SDI scores range from 1 to 100, with higher scores indicating greater neighborhood deprivation.28Scores at the Zip Code Tabulation Area (ZCTA)- level were assigned to individual patients using their residential address.

[0070] Clinical diagnoses were considered present if the International Classification of Diseases, 9thRevision [ICD-9] or ICD-10 code for that specific condition was recorded during hospitalization or in at least 2 outpatient encounters. This methodology has been used extensively in research conducted with VHA and Medicare claims data, and has been shown to enhance diagnostic accuracy in these data sources.29Time-varying predictors included daily physiological parameters obtained from ICDs (e.g., physical activity, atrial fibrillation burden; see eTables 1 a and 1 b for a complete list), arrhythmia episodes, and device therapies (number of ATP and number of shocks delivered). In models predicting death, interim HF hospitalizations were included as time-varying predictors.

[0071] Statistical Analysis

[0072] Categorical variables were reported as frequencies and percentages, and continuous variables were reported as means and standard deviations. We selected random forests for survival, longitudinal and multivariate data (RF-SLAM) algorithm to build the risk prediction model because of its ability to handle time-varying predictors and dynamic effects. RF-SLAM has been used in prior studies to estimate risk of malignant arrhythmias in ICD patients30 31and details of this ML method have been described elsewhere.11Briefly, the RF-SLAM algorithm uses Poisson regression log-likelihood as the split statistic to allow the inclusion of time-varying predictors and relax the proportional hazard assumption by regression methods. The estimated hazard rate is ensembled across trees using Bayesian methodology assuming a Gamma prior distribution. Since the primary goal of the study was to determine which factors were most essential to optimize prediction models, we tested a series of models to evaluate the added value of increasingly complex clinical data: Model 1 time-fixed predictors (baseline demographic and clinical characteristics from the time of implant); Model 2 time-varying predictors (age, sex, CRT therapy, and time-varying device predictors); Model 3 all predictors (baseline demographic and clinical characteristics from the time of implant, time-varying device predictors, time-varying HF hospitalizations, employment status, SDI score).

[0073] First, we preprocessed the time-varying events and covariates for each individual into person-time intervals by partitioning the dataset into 3-month bins to reflect typical remote monitoring follow-up schedules for patients with ICDs / CRT-Ds.24The RF-SLAM approach captures the exact time at which an event or censor occurs within each bin. Thus, the length of risk time within each bin may vary, with measurements recorded up to the date of an event or censor, rather than at the beginning of each bin. For each prediction model, we grew 100 trees, and prespecified the minimum terminal node size (nodesize) as 10% of the total number of person-times and the number of variables tested at each potential split (mtry) as the square root of the number of candidate predictors in the model. Missing data was addressed using an adaptive tree imputation method32(details in eMethods). The impact of individual variables on model performance (i.e., variable importance) was assessed by measuring the percentage of trees that use the variable, and partial dependence plots were used to visualize the relationship between top predictors and the predicted risk. To ensure the clinical relevance of the predicted outcomes, we scaled the risk prediction across the range of risks and over time by using logistic regression with binary events on natural spline functions of the out-of-bag predictions and time-to-event, each with 2 degrees of freedom.

[0074] To characterize the model’s performance, we considered both its discrimination and calibration. Model discrimination was evaluated using timevarying area under the receiver operating curve (AUC). As in other studies,30we also compared the discriminative performance of RF-SLAM to a traditional multivariable time-varying Cox model that included predictors that were used for > 15% of the trees. The best-performing RF-SLAM model in the training dataset was selected for further calibration and external validation in a separate, non-Veteran cohort of ICD patients from a large academic healthcare system. Model calibration was assessed using calibration curve and Cox’s intercept and slope by regressing the observed binary outcome to the log odds of the predicted risk. An intercept equal to 0 and a slope equal to 1 demonstrate good calibration. For the external validation, we obtained the predicted risk by sending person-time observations down the tree, following the branches to the left to right based upon the covariate values and decision rule at each split until reaching the terminal node, and assembling the estimated hazard rate by averaging across all the trees. All statistical analyses were performed using R version 4.1 .3 (R Foundation for Statistical Computing, Austria). Results

[0075] Characteristics of the study cohorts

[0076] A study flow diagram is provided in Figure 1 . Baseline characteristics of the 12,043 patients included in the VHA training cohort and 1 ,397 patients included in the validation cohort are shown in Table 1. The median (IQR) follow-up length was 3.3 (1.9-4.9) years for the training cohort and was 3.6 (2.1 -5.3) years for the validation cohort. There were substantially more women (33% vs 2%) and non-white individuals (32% vs 23%) in the UNC validation cohort, and UNC patients were less likely to be current smokers (10% vs 40%) or have comorbid substance-related disorder (6% vs 11 %) or mental health diagnoses. In the VHA training cohort, 194 (1 .6%) had a prior SCA and 3,478 (28.9%) had an episode of ventricular tachycardia / ventricular fibrillation documented in the EHR prior to ICD implant, suggesting that the majority of the cohort was likely implanted for primary prevention. Data on primary vs. secondary prevention were not available for the UNC cohort. A higher proportion of UNC patients had HF (98% vs 71 %), previous myocardial infarction (81 % vs 17%), and were prescribed guideline-directed therapies for HF and arrythmia management at the time of ICD implant. Device-derived physiological parameters were also generally similar between the two groups (eTables 1a and 1b in the Supplemental Material section below), though the VHA training cohort had a slightly higher burden of AF, HRV, and less percentage of ventricular pacing. The primary outcome of all-cause mortality was observed in 10.6% (n = 1 ,272) of the training cohort and in 8.2% (n = 115) of the validation cohort. The composite mortality / HF hospitalization outcome occurred in 20.5% (n = 2,467) of patients in the training cohort and 25.3% (n = 353) of patients in the validation cohort during the follow up period. Incidence rates for each outcome are reported in eTable 2.

[0077] Variable Importance A total of 47 predictor covariates were incorporated into the risk prediction models. The most important variables for predicting each outcome (measured by the percentage of trees that used the variable) are displayed in Figures 2A-2D. Time-varying device measurements, particularly daily physical activity, consistently emerged as the most important predictors of both outcomes whereas baseline clinical data and static measurements from the time of implant (e.g., LVEF) were less important in predicting outcomes. However, there were notable differences in the variable importance for each outcome. Daily physical activity (59%), daily thoracic impedance (57%), HRV (45%), age (43%), and number of ICD shocks (41 %) were among the strongest predictors of 3-month mortality risk. Similarly, daily physical activity (76%) emerged as the most important predictor of 1 -year mortality risk, while AF burden (47%) and the percentage of ventricular pacing (38%) were also influential predictors. In models predicting the composite outcome of death / HF hospitalization at 3-months, the top predictors were all derived from daily device data: thoracic impedance (65%), physical activity (59%), HRV (55%), and AF burden (41 %). In the model predicting death / HF hospitalization at 1 - year, the most important variables included daily device data [thoracic impedance (67%) and physical activity (61 %)]. Clinical characteristics, including chronic kidney disease (50%) and baseline HF (47%), also emerged as important predictors. The relationships between individual variables and the predicted risks (i.e., direction of associations) are shown in dependence plots in Figures 4-7.

[0078] RF-SLAM Model Performance and Comparison with Traditional Risk Modeling Approaches

[0079] The predictive performance of the RF-SLAM algorithm is visualized in Figure 3 and model AUCs are presented in eTable 3. RF-SLAM Model 1 (baseline predictors) produced an AUC and 95% confidence interval (95% Cl) for predicting 3-month mortality of 0.77 (0.72 - 0.82) and 0.74 (0.72 - 0.76) for 1 year mortality. The discriminative performance of the algorithm improved significantly in Model 2 which included age, sex, CRT, and time-varying device predictors [AUC of 0.91 (0.87 - 0.94) when predicting all-cause mortality at 3 months and 0.80 (0.78-0.82) at 1 - year]. Inclusion of all predictors in Model 3 (baseline clinical and demographic characteristics, interim HF hospitalizations, time-varying predictors, and socioeconomic factors) did not substantially improve model performance, as evidenced by AUCs 0.88 (0.85 - 0.92) at 3-months and 0.80 (0.78 - 0.82) 1 -year. Similarly, RF-SLAM Model 2 emerged as the optimal prediction model for estimating an individual’s risk of death / HF hospitalization at 3 months [0.81 (0.79-0.83)] and 1 -year [ 0.71 (0.70-0.72)].

[0080] Compared to other models (Figure 3), RF-SLAM Model 2 demonstrated better discriminative performance than prediction models using only time-fixed baseline characteristics (RF-SLAM Model 1 ) and traditional regression methods (multivariable time-varying Cox model).

[0081] RF-SLAM Model 2 Performance in Subgroups, Calibration, and External Validation

[0082] Since RF-SLAM Model 2 emerged as the best-performing model in the training dataset, it was selected for further evaluation in predefined subgroups, calibration, and external validation. RF-SLAM Model 2 performance was generally equivalent for men and women, as well as for different categories of race (eTable 4). Calibration intercepts and slopes for RF-SLAM Model 2 are shown in Table 2; model calibration plots are shown in Figures 8-11 in the Data Supplement. RF-SLAM Model 2 (time-varying device predictors) demonstrated excellent calibration and was able to accurately differentiate low and high-risk individuals.

[0083] Next, the performance of RF-SLAM Model 2 was assessed in an external cohort of non-Veteran ICD patients from UNC (eTable 5). The AUC for the risk of death was 0.91 (0.82- 1 .00) at 3-months and 0.79 (0.70 - 0.88) at 1 Year with consistent performance over time [0.80 (0.77 - 0.82) in Years 3-5], There was a slight decline in model performance for predicting death / HF hospitalization in the external validation cohort with lower reported AUCs at 3- months and 1-year. Model discriminations and calibrations for risk of death and death / HF hospitalization in the validation cohort were similar to the training cohort (Table 2, Figure 12).

[0084] Discussion

[0085] In this study, we trained and externally validated a novel ML algorithm that provides clinically meaningful predictions of death from any cause and death / HF hospitalization in patients with ICDs using variables that are readily available to treating clinicians. We evaluated 3 unique ML-based algorithms and found that the model including basic demographic data (i.e., age, sex, CRT therapy) routinely recorded at the time of implant, along with time-varying physiological data that is automatically collected by ICDs, performed best, with excellent discrimination and calibration in the external validation cohort and improved performance compared to traditional risk prediction tools. We also observed that continuous physiologic data from ICDs, particularly daily physical activity, had substantial importance in predicting short-term (3- month) and long-term (1 -year) health risks among patients with ICDs. These findings suggest that dynamic ML models, such as RF-SLAM, could be used to provide early and accurate identification of ICD patients who may benefit from targeted preventive interventions.

[0086] Our study extends previous work by leveraging rigorous ML methods to develop and independently validate a superior risk prediction model that could be applied to a heterogeneous population of ICD patients, readily integrated into clinical practice, easily interpreted by healthcare providers, and potentially shared directly with patients to promote education and risk factor modification. Consistent with these goals, we trained and tested our algorithm in two, large unrelated cohorts (nationwide sample of veterans and non-VHA patients from a large academic health system) and included patients who are typically excluded from clinical trials (adults >75 years; individuals with multiple chronic medical and psychiatric comorbidities), though are often recipients of these devices and may benefit the most from individual risk assessment and timely intervention. Subgroup analyses conducted in both cohorts provide further evidence of the model’s performance in different patient groups and clinical settings. In this regard, our algorithm provides more robust generalizability than previous risk prediction models.

[0087] The RF-SLAM algorithm also improves the discriminative performance and clinical utility of previous prediction models by analyzing complex nonlinear relationships among predictors, capturing dynamic changes in disease trajectories, and automatically generating individual risk predictions that are updated over time. Among the ML models tested, Model 2, which relies on data routinely collected in ICD patients (baseline demographic entered at the time of implant and daily ICD parameters automatically collected by the device and are summarized on remote monitoring reports), provided the best discriminant ability for all-cause mortality and effectively identified individuals at high and low risk of adverse outcomes. It also outperformed conventional risk prediction models evaluated in this study, including models that relied on time-fixed patient characteristics and conventional statistical methods (timevarying Cox models).

[0088] Compared to previously published models, our dynamic ML-based model provided more accurate risk estimates of all-cause mortality (AUC’s = 0.91 at 3 months, 0.80 at 1 -year) than existing models derived from identical device sensor data and less robust statistical methods (AUC =0.72).33Moreover, although we were unable to conduct a direct comparison of our algorithm to two well-accepted risk models for predicting death in defibrillator patients with heart failure [Meta-Analysis Global Group in Chronic Heart Failure (MAGGIC)34and the Seattle Heart Failure Model (SHFM)],35these manually calculated risk scores have consistently demonstrated suboptimal risk discrimination and calibration in prior studies (MAGGIC risk score AUC = 0.74 at 1 year34and SHFM AUC’s = 0.66, 0.67, and 0.68, at 1 , 2, and 5 years).35Population-based risk models, such as the MAGGIC and SHFM, have also been shown to be poor predictors of individual patient outcomes.36 Additionally, while slightly lower AUCs were found for the composite outcome of death / HF hospitalization in this study (AUC’s of 0.81 at 3 months to 0.71 at 1 -year), these results are still comparable or superior to existing prognostic models developed using traditional statistical approaches (O.73)20and ML methods (0.76-0.78).3738These findings highlight the power, performance, and potential scalability of our ML-based prognostic model for predicting adverse outcomes in patients with ICDs.

[0089] It is notable that model discrimination was less robust for the composite outcome of death / HF hospitalization. Similar declines in model performance have been reported for composite outcomes in prior studies39and may reflect the fact that some hospitalizations for HF were missed, particularly in the validation cohort. Such missing events do have the potential to reduce the overall performance of the model. Additionally, baseline differences between the training and validation cohorts might have contributed to this result. Clinical factors known to influence risk of HF hospitalization including prior myocardial infarction, hypertension, HF at the time of ICD implant and comorbid atrial fibrillation, significantly differed between the 2 cohorts.40Furthermore, there were substantially more women (32% vs 2%) and minorities (32% vs 23%) in the UNC validation cohort compared to the training cohort and prior studies have shown higher rates of HF hospitalization among women and non-white individuals relative to other groups.41 42Therefore, it is imperative that future studies assess the robustness of this model prospectively in diverse populations to insure that risk predictions are accurate, equitable, and impactful.

[0090] While ML models, such as RF-SLAM, offer superior risk estimates, they often lack transparency (“black boxes”) and can be difficult for providers to interpret. To make a complex modeling approach more familiar and interpretable to practicing clinicians, we visualized the influence of every single parameter on the predictive performance of the model using intuitively recognized bar graphs and dependence plots. These images also demonstrate the superior performance of continuous data streams (e.g., daily physiologic data from ICDs) in predicting short-term and long-term health risks compared to static measurements obtained at the time of implant (e.g., LVEF).

[0091] Implications for Research and Clinical Practice

[0092] While initial results are promising, prospective studies are needed to validate the model in other populations and settings before it is used to guide clinical decisions. Randomized controlled trials are also needed to determine whether algorithm-directed care improves patient outcomes and reduces healthcare costs. Further research is also required to evaluate optimal implementation strategies (e.g., graphical options to facilitate provider interpretation and communication with patients) and to determine whether sharing individual risk profiles directly with patients would be beneficial for facilitating shared decisions about treatment strategies, care escalation, and risk modification. Additionally, our models were intentionally built to provide clinically meaningful risk estimates for individual patients at 3-month intervals to coincide with the routine review of ICD data from remote monitoring reports. Future studies should evaluate additional risk windows (1 week, 30 days, etc.) to determine whether the RF-SLAM algorithm can be used to improve assessment of acute risk of death or HF decompensation.

[0093] If validated in future studies, our automated algorithm can be easily integrated into existing remote monitoring reports and used by a wide range of providers to identify high-risk individuals who are most likely to benefit from preventative interventions. For example, because our model provides highly accurate and continuously updated estimates of all-cause mortality, these data could be used to discuss prognosis with patients and families, and to facilitate shared decisions regarding treatment (e.g., primary prevention ICD generator replacement). Our algorithm may also help guide the frequency of follow-up (remote and in-person) and initiation or titration of pharmacotherapy. In addition, by identifying unique sources of risk for each patient, our ML algorithm provides an opportunity for personalized patient education and risk modification. In particular, physical activity emerged as an important - potentially modifiable - risk factor for all outcomes in this study. A dynamic risk calculator that receives automated inputs from the RF-SLAM algorithm could be built into a companion website, smartphone app, or embedded directly into remote monitoring reports to visualize an individual’s current risk profile, generate a personalized exercise prescription (e.g., exact number of minutes per day of activity needed to reduce their risk of adverse outcomes), and prompt referrals to exercise training programs (e.g., cardiac rehabilitation).

[0094] Limitations

[0095] This study has several limitations. Due to the retrospective nature of this study and data availability, we were unable to incorporate all factors known to influence prognosis, morbidity, and mortality in patients with ICDs [e.g., device indication (primary vs secondary), heart failure etiology and subtypes, adjustments to medications, LVAD implant or loss of limbs during follow up]. While the AUCs achieved by the RF-SLAM algorithm in our study indicate a very high level of accuracy (e.g., 0.91 ), inclusion of additional laboratory values - including proinflammatory biomarkers43, genetic testing, and imaging data might have improved risk prediction. These data were not available in the data sets used. However, an overarching goal of this study was to create a practical risk prediction tool with existing data that does not require EHR integration, data input, or manual calculation of scores which have been identified as key barriers to implementation of existing cardiovascular risk scores.44Thus, we focused on developing an automated risk prediction tool that is highly accurate but utilizes a minimal subset of covariates for computing a prediction. This parsimonious approach to model development may enhance implementation in clinical practice.

[0096] In addition, HF hospitalizations were not adjudicated, which may have impacted model performance for those outcomes. Similarly, data on the cause of death were unavailable and we were therefore unable to determine whether the models performed better at predicting cardiovascular mortality compared to all-cause mortality. Device therapies were also not adjudicated, and we cannot rule out the possibility of misclassification of clinical variables from the EHR. Differences in the characteristics of the training vs. validation cohorts may have impacted model performance in the validation cohort [proportion of patients implanted for primary vs. secondary prevention; inherited cardiomyopathies (e.g., hypertrophic cardiomyopathy); physical disabilities (leg amputation excluded in training cohort)]. Furthermore, although SDI is a comprehensive and validated measure of neighborhood-level social determinants of health, it is based on ZTCA-level data and may not adequately characterize the resources of individual patients. Finally, the devices used in this study were from a single manufacturer. Although similar sensors are available on devices from other manufacturers, this may affect the generalizability of the model.

[0097] Conclusions

[0098] Our study demonstrates the flexibility, feasibility, and performance of a contemporary machine learning algorithm to guide short and long-term risk prognostication in adults with ICDs using variables that are readily available to treating clinicians. Further prospective studies and randomized trials are needed to validate model performance in other populations and settings, and to evaluate its impact on patient outcomes.

[0099] Table 1. Baseline Characteristics of the Study Cohorts.

[0100]

[0101] CRT device= ICD with Cardiac Resynchronization Therapy. LVEF= Left ventricular ejection fraction at the time of implant / HFrEF = heart failure with reduced ejection fraction (LVEF<35%). HFpEF = heart failure with preserved ejection fraction (LVEF>35%). COPD= Chronic obstructive pulmonary disease. PTSD= posttraumtic stress disorder. ACE-I= Angiotensin- converting enzyme - inhibitors. ARB= Angiotensin II receptor blocker.

[0102] SGLT2= Sodium-glucose cotransporter-2 inhibitors.

[0103] *Values are mean ± SD or n (%). SDI= Social deprivation idex. t870 (7.2%) patients in the VHA cohort and 671 (48.0%) patients in the UNC cohort were missing LVEF.

[0104] Table 2. RF-SLAM Model 2 Calibration.

[0105] Model calibration was assessed using calibration curve and Cox’s intercept and slope by regressing the observed binary outcome to the log odds of the predicted risk. An intercept equal to 0 and a slope equal to 1 demonstrate good calibration.

[0106] Supplemental Material

[0107] Methods

[0108] Handling of missing data Missing data was addressed using an adaptive tree imputation method by using non-missing in-bag data to impute missing values before node splitting. This imputation method ensures individuals with missing data remain in the training set and unbiased predictions for variables with a low to moderate amount of missing data.

[0109] The key steps are as follows:

[0110] 1. For each in-bag observation with s missing value in node h , impute by drawing a random value from the non-missing data within the node h. The node h is then split into two daughter nodes based on the imputed data.

[0111] 2. Reset the imputed values in the daughter nodes to missing.

[0112] 3. Repeat Step 1 and 2 until the tree reaches the terminal node.

[0113] 4. After the initial cycle of growing trees, missing data are imputed using the out-of-bag (OOB) summary values.

[0114] 5. Grow a new forest using the imputed data from Step 4.

[0115] 6. For each observation originally missing the value, draw a random value from the non-missing in-bag data within the same terminal node to re-impute the data.

[0116] 7. Repeat Step 5 and 6 iteratively to grow a new forest using the reimputed data. eTable 1a. Time-varying device data per person-3months (VHA training cohort)

[0117] eTable 1 b. Time-varying device data per person-3months (UNC Training

[0118] Cohort)

[0119] In eTables 1a and 1 b, HR=heart rate, HRV=heart rate variability, AF burden=minutes of atrial fibrillation, VT / VF=ventricular tachycardia, and ventricular fibrillation, NST=non-sustained tachycardia, ATP=antitachycardia pacing, and ICD shocks=implantable cardioverter defibrillator shocks. eTable 2. Incidence rates for each outcome by cohort. eTable 3. RF-SLAM Model AUCs for 3-month and 1-year risk predictions with VHA Training Cohort. AUCs for 3-month and 1 -year risk prediction in the VHA training cohort. Columns indicated the model AUC to predict an event in the next three months or one year during the 1styear, 2ndyear or 3rdto 5thyear of follow-up after implantation. eTable 4. AUCs of RF-SLAM Model 2 for 3-month and 1-year risk prediction by sex and race in the VHA Training Cohort and UNC Validation Cohort. eTable 5. External validation of Model 2 AUCs for 3-month and 1 -year risk prediction with UNC cohort. Model 2: Time-Varying

[0120] Time since implant Device Predictors

[0121] 3-month risk prediction 1 -year risk prediction

[0122] Death

[0123] 1styear 0.91 (0.82-1.00) 0.79 (0.70-0.88)

[0124] 2ndyear 0.86 (0.74-0.97) 0.70 (0.64-0.76)

[0125] 3rd-5thyear 0.88 (0.85, 0.92) 0.80 (0.77-0.82)

[0126] Death or HF Hospitalization

[0127] 1styear 0.72 (0.67-0.78) 0.66 (0.62-0.70)

[0128] 2ndyear 0.64 (0.57-0.70) 0.57 (0.54-0.60)

[0129] 3rd-5thyear 0.73 (0.69-0.77) 0.68 (0.65-0.70)

[0130] RF-SLAM Model 2 AUCs for 3-month and 1-year risk prediction in the UNC validation cohort. Columns indicated the model AUC to predict an event in the next three months or one year during the 1styear, 2ndyear or 3rdto 5thyear of follow-up after implantation.

[0131] Figure 13 is a block diagram illustrating an exemplary system for machine-learning-based prediction of adverse outcomes using measured values of time-varying parameters from an implantable or wearable cardiac device. Referring to Figure 13, a computing platform 1300 includes at least one processor 1302 and memory 1304. Computing platform 1300 also includes a trained machine learning model 1306. In one example, machine learning model 1306 may be the RF-SLAM model modified to accept timevarying values of physiological parameters generated from output of one or more embedded sensors of a wearable or implantable cardiac device and output a value indicating a personalized risk estimate of all-cause mortality and / or a composite event of all-cause mortality and heart failure hospitalization. Once trained, machine learning model 1306 can receive, as input, measured time-varying values of physiological parameters generated from output of one or more sensors of an implantable or wearable cardiac device and generate as output a value indicating a personalized risk estimate of all-cause mortality and / or the composite event of all-cause mortality and heart failure hospitalization for the individual subject. Machine learning model 1306 can also output parameter importance values that indicate the relative importance of each of the input parameters in the prediction. Examples of the importance values that can be output by model 1306 indicated by the lengths of the bars in Figures 2A-2D, which indicate a percentage of the trees in the machine learning model that use each parameter. Machine learning model 1306 may be implemented using computer executable instructions stored in memory 1304 and executed by processor 1302.

[0132] Figure 14 is a flow chart illustrating an exemplary process for machine- learning-based prediction of adverse outcomes using measured time-varying values physiological parameters generated from output of one or more embedded sensors of an implantable or wearable cardiac device. Referring to Figure 14, in step 1400, the process includes receiving, as input to a trained machine learning model, measured time-varying values of physiological parameters generated from output of one or more embedded sensors of a wearable or implantable cardiac device worn by or implanted within an individual subject. For example, a subject, such as a human patient or a pre- clinical subject, such as a mouse, may wear a cardiac measurement device, such as a heart rate monitor or may have an implanted cardiac measurement or assistance device, such as an ICD, a pacemaker, an insertable cardiac monitor (ICM), or a cardiac loop recorder, and the measured values of the parameters generated from the output of the sensors of the device may be provided as input to a trained machine learning model. Examples of the parameter values that may be measured and output by wearable or implantable devices are those whose importance values are displayed by the bars for “Device” in Figures 2A-2D. The machine learning model may be an RF-SLAM model trained using the same parameters generated from wearable or implantable devices worn by or implanted within members of a study cohort.

[0133] In step 1402, the process includes generating, as output from the trained machine learning model, a value indicating a personalized risk estimate of all-cause mortality or a composite event of all-cause mortality or heart failure hospitalization for the individual subject. For example, the trained machine learning model may output a value indicating a personalized risk estimate of all-cause mortality for the individual subject and / or a composite event of all-cause mortality or heart-failure hospitalization for the individual subject. The model may also output parameter importance values indicating importances of the parameters contributing to the value indicating the personalized risk estimate of all-cause mortality or the composite event for the individual subject. The importance values are indicated by the lengths of the bars in Figures 2A-2D. The parameter names are displayed adjacent to each bar.

[0134] In step 1404, the process includes performing, based on the value indicating the personalized risk estimate of all-cause mortality or the composite event, an intervention for the individual subject. For example, a physician or technician, given the value indicating the personalized risk estimate of all-cause mortality, the composite event, and / or the parameter importance values may generate a treatment plan for the individual subject to reduce the subject’s risk of all-cause mortality or heart failure hospitalization. The treatment plan may focus on the parameter importance values for the individual subject. For example, if the parameter importance values indicate that the physical activity measurement is the most important factor contributing to the subject’s all-cause mortality risk score, the treatment plan may be directed to increasing the subject’s physical activity. In another example, the intervention may include maintenance or replacement of the implantable or wearable cardiac device, including battery replacement or charging.

[0135] Applications

[0136] The device-parameter-trained models described herein are beneficial because they not only identify whether a patient is at risk, but also a time period to which the risk score applies (i.e., whether the individual will likely encounter the outcome within the specified time period), and source / factors that are driving risk in an individual patient. These data can be used to titrate diuretics, initiate anticoagulation and modify other pharmacotherapies based on an individual’s risk profile. The risk scores described herein can also be used to facilitate conversations with patients and families about prognosis and care escalation. The risk scores and the importance values described herein can be used to identify patients with high arrhythmia burden who may benefit from ablation, left atrial appendage closure and other cardiovascular procedures. The outputs of the models described herein can be used with patients to make a shared decision about whether or not to undergo a procedure to replace a primary prevention ICD when the patient’s current device is at the end of battery life. If a patient does not have an expected survival of >1year - we do not implant a new device when the battery on the patient’s current device ends (usually 8-12 years after implant). These decisions are currently made by a clinician based on experience / subjective estimate of survival. Our model would provide a more accurate method for assessing the risks and benefits of that procedure. This is an increasingly common issue as older adults live longer and may need device replacement. The models described herein will identify whether risk is driven by physical inactivity. When the models indicate that risk of cardiac death or heart failure hospitalization is driven by physical inactivity, such an indication may prompt a physician to refer the patient to an exercise training program or cardiac rehabilitation program to increase activity, which has been shown to reduce CVD morbidity and mortality. This is just one example of any number of health behavior interventions that could be deployed in response to an individual’s risk profile.

[0137] The disclosure of each of the following references is hereby incorporated herein by reference in its entirety.

[0138] References

[0139] 1. Kremers MS, Hammill SC, Berul Cl, et al. The National ICD Registry Report: version 2.1 including leads and pediatrics for years 2010 and 2011. Heart Rhythm. Apr 2013;10(4):e59-65. doi:10.1016 / j.hrthm.2013.01 .035 2. Santangeli P, Di Biase L, Dello Russo A, et al. Meta-analysis: age and effectiveness of prophylactic implantable cardioverter-defibrillators. Ann Intern Med. Nov 02 2010; 153(9):592-9. doi: 10.7326 / 0003-4819-153-9-201011020- 00009

[0140] 3. Poole JE, Olshansky B, Mark DB, et al. Long-Term Outcomes of

[0141] Implantable Cardioverter-Defibrillator Therapy in the SCD-HeFT. Journal of the American College of Cardiology. 2020; 76(4):405-415. doi:doi: 10.1016 / j.jacc.2020.05.061

[0142] 4. Roger VL. Epidemiology of Heart Failure: A Contemporary Perspective. Circ Res. May 14 2021 ; 128(10): 1421 -1434. doi: 10.1161 / circresaha.121 .3181725. Nevzorov R, Goldenberg I, Konstantino Y, et al. Developing a risk score to predict mortality in the first year after implantable cardioverter defibrillator implantation: Data from the Israeli ICD Registry. J Cardiovasc Electrophysiol. Nov 2018; 29(11 ): 1540- 1547. doi: 10.1111 / jce.137256. Ginder C, Li J, Halperin JL, et al. Predicting Malignant Ventricular Arrhythmias Using Real-Time Remote Monitoring. Journal of the American College of Cardiology. 2023 / 03 / 14 / 2023; 81 (10):949- 961 . doi : https: / / doi . org / 10.1016 / j.jacc.2O22.12.024

[0143] 7. Sammut-Powell C, Taylor JK, Motwani M, Leonard CM, Martin GP, Ahmed FZ. Remotely Monitored Cardiac Implantable Electronic Device Data Predict All-Cause and Cardiovascular Unplanned Hospitalization. J Am Heart Assoc. Aug 16 2022; 11 (16):e024526. doi: 10.1161 / jaha.121 .024526

[0144] 8. Boehmer JP, Hariharan R, Devecchi FG, et al. A Multisensor Algorithm Predicts Heart Failure Events in Patients with Implanted Devices: Results from the MultiSENSE Study. JACC Heart Fail. Mar 2017; 5(3):216-225. doi: 10.1016 / j.jchf.2O16.12.011

[0145] 9. Cowie MR, Sarkar S, Koehler J, et al. Development and validation of an integrated diagnostic algorithm derived from parameters monitored in implantable devices for identifying patients at risk for heart failure hospitalization in an ambulatory setting. European Heart Journal. 2013; 34(31 ):2472-2480. doi: 10.1093 / eurheartj / eht083 10. Whellan DJ, Ousdigian KT, Al-Khatib SM, et al. Combined Heart Failure Device Diagnostics Identify Patients at Higher Risk of Subsequent Heart Failure Hospitalizations. Journal of the American College of Cardiology. 2010; 55(17): 1803-1810. doi:doi:10.1016 / j.jacc.2009.11 .089

[0146] 11 . Wongvibulsin S, Wu KC, Zeger SL. Clinical risk prediction with random forests for survival, longitudinal, and multivariate (RF-SLAM) data analysis. BMC Medical Research Methodology. 2019 / 12 / 31 2019; 20(1 ):1. doi: 10.1186 / s12874-019-0863-0

[0147] 12. Kagiyama N, Shrestha S, Farjo PD, Sengupta PP. Artificial Intelligence: Practical Primer for Clinical Research in Cardiovascular Disease. J Am Heart Assoc. Sep 3 2019; 8(17):e012788. doi:10.1161 / jaha.119.012788

[0148] 13. Moons KG, Altman DG, Reitsma JB, et al. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): explanation and elaboration. Ann Intern Med. Jan 6 2015; 162(1 ): W1 -73. doi: 10.7326 / m 14-0698

[0149] 14. Sung RK, Massie BM, Varosy PD, et al. Long-term electrical survival analysis of Riata and Riata ST silicone leads: National Veterans Affairs experience. Heart Rhythm. Dec 2012; 9(12): 1954-61 . doi: 10.1016 / j.hrthm.2O12.08.006

[0150] 15. Maynard C. Ascertaining Veterans’ Vital Status: VA data sources for mortality ascertainment and cause of death. Database & Methods Cyberseminar Series. 2017

[0151] 16. Rosman L, Lampert R, Sears SF, Burg MM. Measuring Physical Activity with Implanted Cardiac Devices: A Systematic Review. J Am Heart Assoc. May 17 2018; 7(11 )doi: 10.1161 / jaha.118.008663

[0152] 17. Rosman L, Salmoirago-Blotcher E, Mahmood R, et al. Arrhythmia Risk During the 2016 US Presidential Election: The Cost of Stressful Politics. J Am Heart Assoc. Jun 2021 ; 10(11 ):e020559. doi: 10.1161 / jaha.120.020559

[0153] 18. Rosman L, Mazzella AJ, Gehi A, et al. Immediate and long-term effects of the COVID-19 pandemic and lockdown on physical activity in patients with implanted cardiac devices. Pacing Clin Electrophysiol. Jan 2022; 45(1 ):111- 123. doi: 10.1111 / pace.14409

[0154] 19. Dusetzina SB, Tyree S, Meyer AM, Meyer A, Green L, Carpenter WR. Linking Data for Health Services Research: A Framework and Instructional Guide. 2014.

[0155] 20. Angraal S, Mortazavi, JB, Gupta, et al. Machine Learning Prediction of Mortality and Hospitalization in Heart Failure with Preserved Ejection Fraction. JACC Heart Fail. Jan 2020; 8(1 ): 12-21. doi: 10.1016 / j.chf.2O19.06.013

[0156] 21. Bradley J, Schelbert EB, Bonnett LJ, et al. Predicting hospitalisation for heart failure and death in patients with, or at risk of, heart failure before first hospitalisation: a retrospective model development and external validation study. Lancet Digit Health. Jun 2022; 4(6):e445-e454. doi:10.1016 / s2589- 7500(22)00045-0

[0157] 22. Huang H, Turner M, Raju S, et al. Identification of acute decompensated heart failure hospitalizations using administrative data. The American Journal of Cardiology. 2017; 119(11 ): 1791 -1796.

[0158] 23. Bradley J, Schelbert EB, Bonnett LJ, et al. Predicting hospitalisation for heart failure and death in patients with, or at risk of, heart failure before first hospitalisation: a retrospective model development and external validation study. Lancet Digit Health. Jun 2022; 4(6):e445-e454. doi:10.1016 / s2589- 7500(22)00045-0

[0159] 24. Ferrick AM, Raj SR, Deneke T, et al. 2023 HRS / EHRA / APHRS / LAHRS Expert Consensus Statement on Practical Management of the Remote Device Clinic. Europace. May 19 2023; 25(5)doi:10.1093 / europace / euad123

[0160] 25. Hammond G, Johnston K, Huang K, Joynt Maddox KE. Social

[0161] Determinants of Health Improve Predictive Accuracy of Clinical Risk Models for Cardiovascular Hospitalization, Annual Cost, and Death. Circ Cardiovasc Qual Outcomes. Jun 2020; 13(6):e006752. doi: 10.1161 / circoutcomes.120.006752

[0162] 26. Patel SA, Krasnow M, Long K, Shirey T, Dickert N, Morris AA. Excess 30-Day Heart Failure Readmissions and Mortality in Black Patients Increases with Neighborhood Deprivation. Circ Heart Fail. Dec 2020; 13(12):e007947. doi: 10.1161 Zcircheartfailure.120.007947

[0163] 27. Butler DC, Petterson S, Phillips RL, Bazemore AW. Measures of social deprivation that predict health care access and need within a rational area of primary care service delivery. Health Serv Res. Apr 2013; 48(2 Pt 1 ):539-59. doi: 10.1111 / j.1475-6773.2012.01449.x

[0164] 28. Wadhera RK, Bhatt DL, Kind AJH, et al. Association of Outpatient Practice-Level Socioeconomic Disadvantage with Quality of Care and Outcomes Among Older Adults with Coronary Artery Disease: Implications for Value-Based Payment. Circ Cardiovasc Qual Outcomes. Apr 2020; 13(4):e005977. doi: 10.1161 / circoutcomes.119.00597

[0165] 29. Tirschwell DL, Longstreth WT, Jr. Validating administrative data in stroke research. Stroke. Oct 2002; 33(10):2465-7

[0166] 30. Kolk MZH, Ruiperez-Campillo S, Alvarez-Florez L, et al. Dynamic prediction of malignant ventricular arrhythmias using neural networks in patients with an implantable cardioverter-defibrillator. EBioMedicine. Jan 2024; 99:104937. doi: 10.1016 / j.ebiom.2023.104937

[0167] 31. Wu KC, Wongvibulsin S, Tao S, et al. Baseline and Dynamic Risk Predictors of Appropriate Implantable Cardioverter Defibrillator Therapy. J Am Heart Assoc. Oct 20 2020; 9(20):e017002. doi: 10.1161 / jaha.120.017002

[0168] 32. Ishwaran H, Kogalur UB, Blackstone EH, Lauer MS. Random survival forests. The annals of applied statistics. 2008;2(3):841 -860

[0169] 33. Ahmed FZ, Sammut-Powell C, Kwok CS, et al. Remote monitoring data from cardiac implantable electronic devices predicts all-cause mortality. EP Europace. 2021 ; 24(2):245-255. doi:10.1093 / europace / euab160

[0170] 34. Sartipy U, Dahlstrdm U, Edner M, Lund LH. Predicting survival in heart failure: validation of the MAGGIC heart failure risk score in 51 ,043 patients from the Swedish heart failure registry. Eur J Heart Fail. Feb 2014; 16(2): 173- 9. doi: 10.1111 / ejhf.32 35. Regoli F, Scopigni F, Leyva F, et al. Validation of Seattle Heart Failure Model for mortality risk prediction in patients treated with cardiac resynchronization therapy. European Journal of Heart Failure. 2013; 15(2):211 -220. doi : https: / / doi . org / 10.1093 / eurjhf / hfs162

[0171] 36. Allen LA, Matlock DD, Shetterly SM, et al. Use of Risk Models to Predict

[0172] Death in the Next Year Among Individual Ambulatory Patients with Heart Failure. JAMA Cardiology. 2017; 2(4):435-441 . doi: 10.1001 / jamacardio.2016.5036

[0173] 37. Mpanya D, Celik T, Klug E, Ntsinjana H. Predicting mortality and hospitalization in heart failure using machine learning: A systematic literature review. Int J Cardiol Heart Vase. Jun 2021 ; 34:100773. doi: 10.1016 / j. ijcha.2021 .100773

[0174] 38. Desai RJ, Wang SV, Vaduganathan M, Evers T, Schneeweiss S.

[0175] Comparison of Machine Learning Methods with Traditional Models for Use of Administrative Claims with Electronic Medical Records to Predict Heart Failure Outcomes. JAMA Netw Open. Jan 3 2020; 3(1 ):e1918962. doi: 10.1001 / jamanetworkopen.2019.18962

[0176] 39. D’Onofrio A, Solimene F, Calo L, et al. Combining home monitoring temporal trends from implanted defibrillators and baseline patient risk profile to predict heart failure hospitalizations: results from the SELENE HF study. EP Europace. 2021 ; 24(2):234-244. doi: 10.1093 / europace / euab170

[0177] 40. Banerjee A, Pasea L, Chung SC, et al. A population-based study of 92 clinically recognized risk factors for heart failure: co-occurrence, prognosis and preventive potential. Eur J Heart Fail. Mar 2022; 24(3):466-480. doi:10.1002 / ejhf.2417

[0178] 41. Keshvani N, Shah S, Ayodele I, et al. Sex differences in long-term outcomes following acute heart failure hospitalization: Findings from the Get with The Guidelines-Heart Failure Registry. Eur J Heart Fail. Sep 2023; 25(9): 1544-1554. doi: 10.1002 / ejhf.3003 42. Pina IL, Jimenez S, Lewis EF, et al. Race and Ethnicity in Heart Failure: JACC Focus Seminar 8 / 9. Journal of the American College of Cardiology. 2021 / 12 / 21 / 2021 ;78(25):2589-2598. doi: https: / / doi . org / 10.1016 / j.jacc.2O21 .06.058 43. Segar MW, Jaeger BC, Patel KV, et al. Development and Validation of

[0179] Machine Learning-Based Race-Specific Models to Predict 10-Year Risk of Heart Failure: A Multicohort Analysis. Circulation. Jun 152021 ; 143(24):2370- 2383. doi: 10.1161 / circulationaha.120.053134

[0180] 44. Tuzzio L, O'Meara ES, Holden E, et al. Barriers to Implementing Cardiovascular Risk Calculation in Primary Care: Alignment with the Consolidated Framework for Implementation Research. Am J Prev Med. Feb 2021 ; 60(2):250-257. doi:10.1016 / j.amepre.2020.07.027

[0181] It will be understood that various details of the subject matter described herein may be changed without departing from the scope of the subject matter described herein. Furthermore, the foregoing description is for the purpose of illustration only, and not for the purpose of limitation, as the subject matter described herein is defined by the claims as set forth hereinafter.

Claims

1. CLAIMSWhat is claimed is:1 . A method for machine-learning-based prediction of adverse outcomes using measured time-varying values of physiological parameters generated from output of one or more embedded sensors of an implantable or wearable cardiac device, the method comprising: receiving, as input to a trained machine learning model, measured time-varying values of physiological parameters generated from output of a wearable or implantable cardiac device worn by or implanted within an individual subject; generating, as output from the trained machine learning model, a value indicating a personalized risk estimate of all-cause mortality or a composite event of all-cause mortality or heart failure hospitalization for the individual subject; and performing, based on the value indicating the personalized risk estimate of all-cause mortality or the composite event, an intervention for the individual subject.

2. The method of claim 1 wherein the wearable or implantable cardiac device comprises a cardiovascular implantable electronic device (CIED).

3. The method of claim 2 wherein the CIED comprises an implantable cardioverter defibrillator (ICD) with or without cardiac resynchronization therapy (CRT-D), an insertable cardiac monitor (ICM), a cardiac loop recorder, or a pacemaker with or without cardiac resynchronization therapy (CRT-P).

4. The method of claim 1 wherein the wearable or implantable cardiac device comprises a wearable device.

5. The method of claim 4 wherein the wearable device comprises a wearable heart rate monitor.

6. The method of claim 4 wherein the wearable device comprises a smart watch or a smart ring.

7. The method of claim 1 comprising, outputting, by the trained machine learning model, parameter importance values indicating importances of the physiological parameters contributing to the value indicating the personalized risk estimate of all-cause mortality or the composite event for the individual subject.

8. The method of claim 1 wherein the value indicating the personalized risk estimate of all-cause mortality or the composite event of all-cause mortality or heart failure hospitalization for the individual subject is associated with a time period when the all-cause mortality or a composite event of all-cause mortality or heart failure hospitalization is likely to occur.

9. The method of claim 1 wherein the machine learning model comprises a random forest for survival, longitudinal and multivariate (RF-SLAM) model.

10. The method of claim 9 wherein the RF-SLAM model is trained using measured time-varying values of physiological parameters generated from output of one or more embedded sensors of wearable or implantable cardiac devices worn by or implanted within subjects in a study cohort.

11. The method of claim 1 wherein performing the intervention includes performing replacement, repair, or maintenance of the wearable or implantable cardiac device.

12. The method of claim 1 wherein performing the intervention includes generating a cardiac treatment plan tailored to reduce the value indicating the personalized risk estimate of all-cause mortality or the composite event.

13. A system for machine-learning-based prediction of adverse outcomes using measured time-varying values of physiological parametersgenerated from output of one or more embedded sensors an implantable or wearable cardiac device, the system comprising: a computing platform including at least one processor and a memory; and a trained machine learning model comprising computerexecutable instructions stored in the memory and executable by the at least one processor for receiving, as input, measured time-varying values of the physiological parameters generated from output of one or more embedded sensors of a wearable or implantable cardiac device worn by or implanted within an individual subject and generating, as output, a value indicating a personalized risk estimate of all-cause mortality or a composite event of all-cause mortality or heart failure hospitalization for the individual subject.

14. The system of claim 13 wherein the wearable or implantable cardiac device comprises a cardiac implantable electronic device (CIED).

15. The system of claim 14 wherein the CIED comprises an implantable cardioverter defibrillator (ICD) with or without cardiac resynchronization therapy (CRT-D), an insertable cardiac monitor (ICM), a coronary pressure-sensing device, a cardiac loop recorder, or a pacemaker with or without cardiac resynchronization therapy (CRT-P).

16. The system of claim 13 wherein the wearable or implantable cardiac device comprises a wearable device.

17. The system of claim 16 wherein the wearable device comprises a wearable heart rate monitor or a smart watch or a smart ring.

18. The system of claim 13 wherein the trained machine learning model is configured to generate, as output, parameter importance values indicating importances of the physiological parameters contributing to the value indicating the personalized risk estimate of all-cause mortality or the composite event for the individual subject.

19. The system of claim 13 wherein the value indicating the personalized risk estimate of all-cause mortality or a composite event of all-causemortality or heart failure hospitalization for the individual subject is associated with a time period when the all-cause mortality or a composite event of all-cause mortality or heart failure hospitalization is likely to occur.

20. The system of claim 13 wherein the trained machine learning model comprises a random forest for survival, longitudinal and multivariate (RF-SLAM) model.

21. The system of claim 20 wherein the RF-SLAM model is trained using measured time-varying values of physiological parameters generated from output of one or more embedded sensors of wearable or implantable cardiac device parameters worn by or implanted within subjects in a study cohort.

22. A non-transitory computer readable medium having stored thereon executable instructions that when executed by a processor of a computer controls the computer to perform steps comprising: receiving, as input to a trained machine learning model, measured time-varying values of physiological parameters generated from output of one or more embedded sensors of a wearable or implantable cardiac device worn by or implanted within an individual subject; and generating, as output from the trained machine learning model, a value indicating a personalized risk estimate of all-cause mortality or a composite event of all-cause mortality or heart failure hospitalization for the individual subject.

Images