The majority of biomedical studies use limited datasets that may not generalize over large heterogeneous datasets that have been collected over several decades. The current paper develops and validates several multimodal models that can predict 1-year mortality based on a massive clinical dataset. Our focus on predicting 1-year mortality can provide a sense of urgency to the patients. Using the largest dataset of its kind, the paper considers the development and validation of multimodal models based on 25,137,015 videos associated with 699,822 echocardiography studies from 316,125 patients, and 2,922,990 8-lead electrocardiogram (ECG) traces from 631,353 patients. Our models allow us to assess the contribution of individual factors and modalities to the overall risk. Our approach allows us to develop extremely low-parameter models that use optimized feature selection based on feature importance. Based on available clinical information, we construct a family of models that are made available in the DISIML package. Overall, performance ranges from an AUC of 0.72 with just ten parameters to an AUC of 0.89 with under 105k for the full multimodal model. The proposed approach represents a modular neural network framework that can provide insights into global risk trends and guide therapies for reducing mortality risk.

Background: Several large trials have employed age or clinical features to select patients for atrial fibrillation (AF) screening to reduce strokes. We hypothesized that a machine learning (ML) model trained to predict AF risk from 12-lead electrocardiogram (ECG) would be more efficient than criteria based on clinical variables in indicating a pop-ulation for AF screening to potentially prevent AF-related stroke.Methods: We retrospectively included all patients with clinical encounters in Geisinger without a prior history of AF. Incidence of AF within 1 year and AF-related strokes within 3 years of the encounter were identified. AF-related stroke was defined as a stroke where AF was diagnosed at the time of stroke or within a year after the stroke. The efficiency of five methods was evaluated for selecting a cohort for AF screening. The methods were selected from four clinical trials (mSToPS, GUARD-AF, SCREEN-AF and STROKESTOP) and the ECG-based ML model. We simulated patient selection for the five methods between the years 2011 and 2014 and evaluated outcomes for 1 year intervals between 2012 and 2015, resulting in a total of twenty 1-year periods. Patients were considered eligible if they met the criteria before the start of the given 1-year period or within that period. The primary outcomes were numbers needed to screen (NNS) for AF and AF-associated stroke.Results: The clinical trial models indicated large proportions of the population with a prior ECG for AF screening (up to 31%), coinciding with NNS ranging from 14 to 18 for AF and 249-359 for AF-associated stroke. At comparable sensitivity, the ECG ML model indicated a modest number of patients for screening (14%) and had the highest efficiency in NNS for AF (7.3; up to 60% reduction) and AF-associated stroke (223; up to 38% reduction).Conclusions: An ECG-based ML risk prediction model is more efficient than contemporary AF-screening criteria based on age alone or age and clinical features at indicating a population for AF screening to potentially prevent AF -related strokes.

Introduction: A deep learning ECG algorithm, rECHOmmend, can accurately identify patients with any of seven structural heart diseases: five valvular diseases, low ejection fraction and interventricular septal (IVS) thickening. Components of the rECHOmmend composite label (IVS>15mm, mitral regurgitation) are also associated with hypertrophic cardiomyopathy (HCM). We hypothesized that despite being trained without HCM-specific labels, rECHOmmend can reliably identify HCM patients and achieve comparable performance to an HCM-specific classifier. Methods: Algorithms were developed from 2,898,979 ECGs acquired from 661,366 patients between 1984-2021. rECHOmmend was trained on a composite label derived from echocardiography and electronic health record (EHR) data. This ensemble model consists of 7 disease specific models with an aggregate model to predict a composite structural heart disease endpoint with shared clinical actionability. Separately, an HCM-specific model was trained on a binary label derived from EHR. To enable comparison, both classifiers were tested on a shared ECG holdout set (ECG prevalence 1.24%, patient prevalence 0.52%). Results: Despite being trained without HCM specific labels, the rECHOmmend ensemble showed comparable performance to a HCM-specific classifier (C-statistic: 0.92 [0.90-0.93] vs 0.90 [0.89-0.91]). At an operating point optimized for the F1-score, the sensitivity to HCM was higher for rECHOmmend at 0.42 [0.33-0.50] compared to 0.18 [0.15-0.21] for the HCM-specific classifier. rECHOmmend sustained performance across a range of IVS thicknesses, suggesting it was not solely reliant on IVS thickening for HCM identification and other ensemble components contributed to performance. Conclusions: A composite deep learning algorithm trained to identify structural heart diseases can identify clinically ascertained HCM with good performance, despite being trained without HCM-specific labels.

Introduction: Cardiac amyloidosis (CA) is a common cause of progressive heart failure. New therapies can improve outcomes but most CA patients remain undiagnosed and untreated. Machine learning models deployed on electronic health record (EHR) data may be able to find patients with undiagnosed CA. To date, most models have focused on identification of undiagnosed amyloid from uncensored data modalities (Fig 1). Hypothesis: We hypothesized that lack of post-diagnosis censoring when training CA models leads to poor performance in predicting patients with undiagnosed CA whereas training with appropriate time censoring improves performance. Methods: We used 41 EHR features (demographics, labs, electrocardiogram/echocardiography measurements, vitals) to train a boosted decision tree model with and without time censoring. This was applied to 112 patients with confirmed CA and 22,400 controls matched on age, sex, encounter frequency and timespan of EHR. We also compared our findings to a web-based CA algorithm that was publicly available in 2020. Results: The EHR algorithm had modestly higher performance on at-risk, time-censored patients when trained with and without time censoring (area under the receiver operating characteristic curve (AUROC) 0.84±0.09 vs 0.79±0.07). Testing on temporally uncensored data showed higher performance (AUROC: 0.91±0.05) which may be unrepresentative of deployment scenarios where post-diagnostic features are unavailable for model use. The publicly available algorithm demonstrated a similar trend when tested on uncensored data (AUROC: 0.67±0.03) as compared to an appropriately censored feature set (AUROC: 0.54±0.04). Conclusions: EHR algorithms can be trained to find patients with high risk of undiagnosed cardiac amyloidosis. These models should be evaluated on temporally censored data so that post-diagnostic features do not artificially inflate performance estimates and negatively impact real-world deployment.

Introduction: Patients at high risk of undiagnosed Atrial Fibrillation (AF) can be identified from 12-lead ECG and routed to increased monitoring via wearable devices such as a watch or an ECG patch. A similar prediction task may be feasible employing echocardiography, increasing the likelihood of detecting patients at high risk for AF. Hypothesis: We hypothesized that by leveraging a large echocardiography database, a model can be trained to accurately predict future AF onset in patients undergoing routine echocardiography. Methods: We collected data on patients with an echo at Geisinger (769,854 studies from 337,388 patients) and randomly grouped patients into training (80%) or testing (20%) sets. In the training cohort, studies were considered positive if they were obtained from patients with prior or new (<1 year) AF diagnosis. Alternatively, negative studies were those obtained from AF-free patients (>1 year). For testing, we selected one random echo per patient and restricted the positives to patients without prior AF diagnosis or new AF within 7 days of the echo. This was to avoid predicting positive AF at the time of echo. We tested the model on 38,808 valid studies where 5% developed AF within a year. We trained six convolutional neural networks, one per view (apical two, four and five chambers, basal and mid short-axis, and parasternal long-axis), and combined their outputs to train an XGboost model. Results: Among the six views, the apical four chamber yielded the best performance with an area under the ROC curve (AUROC) of 76%, followed by the apical two chamber with 65%, the apical five chamber with 58%, the parasternal long-axis with 57%, the basal short-axis with 56%, and the mid short-axis with 50%. A model combining all six planes yielded a 78% AUROC. Conclusions: CNN models trained on echocardiography videos can predict new-onset AF with moderate performance. We also found that views that captured the left atrium had better performance compared to those that did not.

BACKGROUND:Timely diagnosis of structural heart disease improves patient outcomes, yet many remain underdiagnosed. While population screening with echocardiography is impractical, ECG-based prediction models can help target high-risk patients. We developed a novel ECG-based machine learning approach to predict multiple structural heart conditions, hypothesizing that a composite model would yield higher prevalence and positive predictive values to facilitate meaningful recommendations for echocardiography. METHODS:Using 2 232 130 ECGs linked to electronic health records and echocardiography reports from 484 765 adults between 1984 to 2021, we trained machine learning models to predict the presence or absence of any of 7 echocardiography-confirmed diseases within 1 year. This composite label included the following: moderate or severe valvular disease (aortic/mitral stenosis or regurgitation, tricuspid regurgitation), reduced ejection fraction <50%, or interventricular septal thickness >15 mm. We tested various combinations of input features (demographics, laboratory values, structured ECG data, ECG traces) and evaluated model performance using 5-fold cross-validation, multisite validation trained on 1 site and tested on 10 independent sites, and simulated retrospective deployment trained on pre-2010 data and deployed in 2010. RESULTS:Our composite rECHOmmend model used age, sex, and ECG traces and had a 0.91 area under the receiver operating characteristic curve and a 42% positive predictive value at 90% sensitivity, with a composite label prevalence of 17.9%. Individual disease models had area under the receiver operating characteristic curves from 0.86 to 0.93 and lower positive predictive values from 1% to 31%. Area under the receiver operating characteristic curves for models using different input features ranged from 0.80 to 0.93, increasing with additional features. Multisite validation showed similar results to cross-validation, with an aggregate area under the receiver operating characteristic curve of 0.91 across our independent test set of 10 clinical sites after training on a separate site. Our simulated retrospective deployment showed that for ECGs acquired in patients without preexisting structural heart disease in the year 2010, 11% were classified as high risk and 41% (4.5% of total patients) developed true echocardiography-confirmed disease within 1 year. CONCLUSIONS:An ECG-based machine learning model using a composite end point can identify a high-risk population for having undiagnosed, clinically significant structural heart disease while outperforming single-disease models and improving practical utility with higher positive predictive values. This approach can facilitate targeted screening with echocardiography to improve underdiagnosis of structural heart disease.

Use of machine learning (ML) for automated annotation of heart structures from echocardiographic videos is an active research area, but understanding of comparative, generalizable performance among models is lacking. This study aimed to (1) assess the generalizability of five state-of-the-art ML-based echocardiography segmentation models within a large Geisinger clinical dataset, and (2) test the hypothesis that a quality control (QC) method based on segmentation uncertainty can further improve segmentation results. Five models were applied to 47,431 echocardiography studies that were independent from any training samples. Chamber volume and mass from model segmentations were compared to clinically-reported values. The median absolute errors (MAE) in left ventricular (LV) volumes and ejection fraction exhibited by all five models were comparable to reported inter-observer errors (IOE). MAE for left atrial volume and LV mass were similarly favorable to respective IOE for models trained for those tasks. A single model consistently exhibited the lowest MAE in all five clinically-reported measures. We leveraged the tenfold cross-validation training scheme of this best-performing model to quantify segmentation uncertainty. We observed that removing segmentations with high uncertainty from 14 to 71% studies reduced volume/mass MAE by 6-10%. The addition of convexity filters improved specificity, efficiently removing < 10% studies with large MAE (16-40%). In conclusion, five previously published echocardiography segmentation models generalized to a large, independent clinical dataset-segmenting one or multiple cardiac structures with overall accuracy comparable to manual analyses-with variable performance. Convexity-reinforced uncertainty QC efficiently improved segmentation performance and may further facilitate the translation of such models.

Machine learning promises to assist physicians with predictions of mortality and of other future clinical events by learning complex patterns from historical data, such as longitudinal electronic health records. Here we show that a convolutional neural network trained on raw pixel data in 812,278 echocardiographic videos from 34,362 individuals provides superior predictions of one-year all-cause mortality. The model's predictions outperformed the widely used pooled cohort equations, the Seattle Heart Failure score (measured in an independent dataset of 2,404 patients with heart failure who underwent 3,384 echocardiograms), and a machine learning model involving 58 human-derived variables from echocardiograms and 100 clinical variables derived from electronic health records. We also show that cardiologists assisted by the model substantially improved the sensitivity of their predictions of one-year all-cause mortality by 13% while maintaining prediction specificity. Large unstructured datasets may enable deep learning to improve a wide range of clinical prediction models.

Background Severity scores assess the acuity of critical illness by penalizing for the deviation of physiologic measurements from normal and aggregating these penalties (also called “weights” or “subscores”) into a final score (or probability) for quantifying the severity of critical illness (or the likelihood of in-hospital mortality). Although these simple additive models are human readable and interpretable, their predictive performance needs to be further improved. Methods We present OASIS +, a variant of the Oxford Acute Severity of Illness Score (OASIS) in which an ensemble of 200 decision trees is used to predict in-hospital mortality based on the 10 same clinical variables in OASIS. Results Using a test set of 9566 admissions extracted from the MIMIC-III database, we show that OASIS + outperforms nine previously developed severity scoring methods (including OASIS) in predicting in-hospital mortality. Furthermore, our results show that the supervised learning algorithms considered in our experiments demonstrated higher predictive performance when trained using the observed clinical variables as opposed to OASIS subscores. Conclusions Our results suggest that there is room for improving the prognostic accuracy of the OASIS severity scores by replacing the simple linear additive scoring function with more sophisticated non-linear machine learning models such as RF and XGB.

Background: Several large trials have employed age or clinical features to select patients for atrial fibrillation (AF) screening to reduce strokes. We hypothesized that a deep neural network (DNN) model risk prediction based on ECG would be superior to age and clinical variables at selecting a population at high risk for AF and AF-related stroke. Methods: We retrospectively included all patients with an ECG at Geisinger without a prior history of AF. Incidence of AF and AF-related strokes were identified as outcomes within 1 and 3 years after the ECG, respectively. AF-related stroke was defined as a stroke where AF was diagnosed at the time of stroke or within a year after the stroke. We selected a high-risk cohort for AF screening based on five risk stratification methods - criteria from four clinical trials (mSToPS, STROKESTOP, GUARD-AF and SCREEN-AF) and the DNN model at the qualifying ECG. We simulated patient selection and evaluated outcomes for twenty 1-year periods between 2010-2014 centered around the ECG encounter. For the clinical trials, the patients were considered eligible if they met the criteria before or within the period unless they satisfied exclusion criteria at the time of ECG. Results: The DNN model achieved optimal sensitivity (65%), PPV (10%), NNS for AF (10) within this population compared with all other risk models with a NNS for AF-related stroke of 160. Total screening number, sensitivity, positive predictive value (PPV) and number needed to screen (NNS) to capture AF and AF-related stroke are summarized in Table 1. The number of additional screens for the DNN model was slightly higher for two of the other models (SCREEN-AF and STROKESTOP) but lower than the other two (mSToPS and GUARD-AF). Conclusions: A DNN ECG-based risk prediction model is superior to contemporary AF-screening criteria based on age alone or age and clinical features in selecting a population for additional screening due to high risk for future AF and potential AF-related strokes.

Background: Atrial fibrillation (AF) is associated with substantial morbidity, especially when it goes undetected. If new-onset AF could be predicted, targeted screening could be used to find it early. We hypothesized that a deep neural network could predict new-onset AF from the resting 12-lead ECG and that this prediction may help identify those at risk of AF-related stroke. Methods: We used 1.6 M resting 12-lead digital ECG traces from 430 000 patients collected from 1984 to 2019. Deep neural networks were trained to predict new-onset AF (within 1 year) in patients without a history of AF. Performance was evaluated using areas under the receiver operating characteristic curve and precision-recall curve. We performed an incidence-free survival analysis for a period of 30 years following the ECG stratified by model predictions. To simulate real-world deployment, we trained a separate model using all ECGs before 2010 and evaluated model performance on a test set of ECGs from 2010 through 2014 that were linked to our stroke registry. We identified the patients at risk for AF-related stroke among those predicted to be high risk for AF by the model at different prediction thresholds. Results: The area under the receiver operating characteristic curve and area under the precision-recall curve were 0.85 and 0.22, respectively, for predicting new-onset AF within 1 year of an ECG. The hazard ratio for the predicted high- versus low-risk groups over a 30-year span was 7.2 (95% CI, 6.9–7.6). In a simulated deployment scenario, the model predicted new-onset AF at 1 year with a sensitivity of 69% and specificity of 81%. The number needed to screen to find 1 new case of AF was 9. This model predicted patients at high risk for new-onset AF in 62% of all patients who experienced an AF-related stroke within 3 years of the index ECG. Conclusions: Deep learning can predict new-onset AF from the 12-lead ECG in patients with no previous history of AF. This prediction may help identify patients at risk for AF-related strokes.

Additional file 1. Supplementary Tables S1–S4.

Introduction: The use of convolutional neural networks (CNN) to automatically segment the heart from echocardiography images has garnered recent attention, but generalizable performance segmenting multiple cardiac structures has not been demonstrated. The objective of the present work was to develop a 2D CNN model to automatically and accurately segment the left ventricular (LV) endo- and epicardial and left atrial (LA) endocardial surfaces from an independent, external clinical dataset. Methods: A modified U-net CNN was trained using 10-fold cross-validation and augmentation within the published CAMUS echocardiography dataset to segment the LV and LA from apical 2- and 4-chamber images. For external validation, this model was applied to 3,087 echocardiograms from Geisinger for which the physician-reported LV ejection fraction (EF) was within 10% of the reported bi-plane EF, signifying confidence in quality of the underlying images. LV end-diastolic and end-systolic volumes (EDV, ESV), LA volume (LAV), and LV mass (LVM) were estimated using Simpson’s bi-plane summation. We compared performance against two published segmentation models. Results: Our model agreed well with clinically-reported values, based on small median absolute errors in percent of clinical measures (MAE; Table 1) and biases, and narrow limits of agreement (mean bias/coefficient of variation (%): LV EDV 10.3/10.6, ESV 6.3/13.1, EF 0.5/6.3, LVM 13.9/10.1, and LAV 3.1/9.5). Moreover, the observed MAE for each metric was within the previously reported limits of inter-observer variability for 2D echo. Compared to two previously published models, our model exhibited smaller MAE in all measures tested (Table 1). Conclusions: Our model exhibits accurate, generalizable performance in multi-structural echocardiography segmentation, with accuracy meeting or exceeding current leading models. Such models hold great promise for translational research and precision medicine efforts.

The electrocardiogram (ECG) is a widely used medical test, consisting of voltage versus time traces collected from surface recordings over the heart1. Here we hypothesized that a deep neural network (DNN) can predict an important future clinical event, 1-year all-cause mortality, from ECG voltage–time traces. By using ECGs collected over a 34-year period in a large regional health system, we trained a DNN with 1,169,662 12-lead resting ECGs obtained from 253,397 patients, in which 99,371 events occurred. The model achieved an area under the curve (AUC) of 0.88 on a held-out test set of 168,914 patients, in which 14,207 events occurred. Even within the large subset of patients (n = 45,285) with ECGs interpreted as ‘normal’ by a physician, the performance of the model in predicting 1-year mortality remained high (AUC = 0.85). A blinded survey of cardiologists demonstrated that many of the discriminating features of these normal ECGs were not apparent to expert reviewers. Finally, a Cox proportional-hazard model revealed a hazard ratio of 9.5 (P < 0.005) for the two predicted groups (dead versus alive 1 year after ECG) over a 25-year follow-up period. These results show that deep learning can add substantial prognostic information to the interpretation of 12-lead resting ECGs, even in cases that are interpreted as normal by physicians. By using data from electrocardiograms, a deep learning algorithm outperforms traditional risk scores in predicting death over the course of the next year and identifies at-risk individuals with seemingly normal electrocardiograms.

AbstractBackgroundAtrial fibrillation (AF) is associated with substantial morbidity, especially when it goes undetected. If new onset AF could be predicted, targeted population screening could be used to find it early. We hypothesized that a deep neural network could predict new onset AF from the resting 12-lead electrocardiogram (ECG) and that this prediction may help prevent AF-related stroke.MethodsWe used 1.6M resting 12-lead ECG voltage-time traces from 430k patients collected from 1984-2019 in this study. Deep neural networks were trained to predict new onset AF (within 1 year) in patients without a history of AF. Performance was evaluated using areas under the receiver operating characteristic curve (AUROC) and precision-recall curve (AUPRC). We performed an incidence-free survival analysis for a period of 30 years following the ECG stratified by model predictions. To simulate real-world deployment, we trained a separate model using all ECGs prior to 2010 and evaluated model performance on a test set of ECGs from 2010 through 2014 that were linked to our stroke registry. We used standard metrics to explore different prediction thresholds for the model and also calculated how many AF-related strokes might be potentially prevented.ResultsThe AUROC and AUPRC were 0.83 and 0.21, respectively, for predicting new onset AF within 1 year of an ECG. Adding age and sex improved the AUROC to 0.85 and the AUPRC to 0.23. The hazard ratio for the predicted high- vs. low-risk groups over a 30-year span was 7.2 [95% confidence interval: 6.9 – 7.6]. In a simulated deployment scenario, using the F2 score to select the risk prediction threshold, the model predicted new onset AF at 1 year with a sensitivity of 69%, specificity of 81%, and positive predictive value (PPV) of 12%. This model correctly predicted new onset AF in 62% of all patients that experienced an AF-related stroke within 3 years of the ECG.ConclusionsDeep learning can predict new onset AF from the 12-lead ECG in patients with no prior history of AF. This prediction may prove useful in preventing AF-related strokes.

BACKGROUND Heart failure is a prevalent, costly disease for which new value-based payment models demand optimized population management strategies. OBJECTIVES This study sought to generate a strategy for managing populations of patients with heart failure by leveraging large clinical datasets and machine learning. METHODS Geisinger electronic health record data were used to train machine learning models to predict 1-year all-cause mortality in 26,971 patients with heart failure who underwent 276,819 clinical episodes. There were 26 clinical variables (demographics, laboratory test results, medications), 90 diagnostic codes, 41 electrocardiogram measurements and patterns, 44 echocardiographic measurements, and 8 evidence-based "care gaps": flu vaccine, blood pressure of <130/80 mm Hg, A(1)c of <8%, cardiac resynchronization therapy, and active medications (active angiotensin-converting enzyme inhibitor/angiotensin II receptor blocker/angiotensin receptor-neprilysin inhibitor, aldosterone receptor antagonist, hydralazine, and evidence-based beta-blocker) were collected. Care gaps represented actionable variables for which associations with all-cause mortality were modeled from retrospective data and then used to predict the benefit of prospective interventions in 13,238 currently living patients. RESULTS Machine learning models achieved areas under the receiver-operating characteristic curve (AUCs) of 0.74 to 0.77 in a split-by-year training/test scheme, with the nonlinear XGBoost model (AUC: 0.77) outperforming linear logistic regression (AUC: 0.74). Out of 13,238 currently living patients, 2,844 were predicted to die within a year, and closing all care gaps was predicted to save 231 of these lives. Prioritizing patients for intervention by using the predicted reduction in 1-year mortality risk outperformed all other priority rankings (e.g., random selection or Seattle Heart Failure risk score). CONCLUSIONS Machine learning can be used to priority-rank patients most likely to benefit from interventions to optimize evidence-based therapies. This approach may prove useful for optimizing heart failure population health management teams within value-based payment models. (c) 2020 by the American College of Cardiology Foundation.

Background: Atrial fibrillation (AF) is associated with stroke, especially when AF goes undetected. Deep neural networks (DNN) can predict incident AF from a 12-lead resting ECG. We hypothesize that use of a DNN to predict new onset AF from an ECG may identify patients at risk of sustaining a potentially preventable AF-related stroke. Methods: We trained a DNN model to predict new-onset AF using 382,604 ECGs prior to 2010. We then evaluated the model performance on a test set of ECGs from 2010 through 2014 linked to patients in an institutional stroke registry. There were 181,969 patients in the test set with at least one ECG and no prior history of AF. Of those patients 3,497 (1.9%) had a stroke following an ECG that did not show AF. Within the set of patients with stroke, 375 had the stroke within 3 years of the ECG and were diagnosed with new AF between -3 and 365 days of the stroke. We considered these potentially preventable AF-related strokes. We report the sensitivity and positive predictive value (PPV) of the model for appropriately risk stratifying these 375 patients who sustained a potentially preventable AF-related stroke. Results: We used F β scores to identify different risk prediction thresholds (operating points) for the model. Operating points chosen by F 0.5 , F 1 , and F 2 scores identified 4, 12, and 21% of the population as high risk for the development of AF within 1 year (Figure 1). Screening 1, 4, 12, and 21% of the overall population resulted in PPV of 28, 21, 15, and 12%, respectively, for identification of new onset AF in one year. Using those same thresholds yielded sensitivities of 4, 17, 45, and 62% for identifying potentially preventable AF-related strokes. The different risk prediction thresholds resulted in a low (120-162) number needed to screen to detect one potentially preventable AF-related stroke at 3 years. Conclusions: Use of a deep learning model to predict new onset AF may identify patients at high risk of sustaining a potentially preventable AF-related stroke.

Introduction: Healthcare cost has increased drastically in the last decade, and over 50% of the cost can be attributed to a small portion (5-10%) of the population. Certain clinical programs, such as home-based care, aim to reduce this utilization but need methods to identify the most appropriate patients to enroll. We hypothesized that machine learning can predict patients with high future utilization with good accuracy. Methods: 683,160 cardiology patients (defined broadly as those with an ECG, echocardiogram or cardiology visit) with ~17 million clinical episodes since 2004 were identified from Geisinger’s electronic health records. Utilization was estimated as total cost of care for outpatient, inpatient and emergency department visits. Patients with the highest 10% utilization in a given year were defined as high utilizers. Machine learning models were used to predict high utilization over the next 3, 6 and 12 months. Input variables (n=191) included age, sex, smoking, 5 vital signs, 21 labs, 18 medications (current and past), 40 ECG and 44 echocardiographic measurements, 43 comorbidities, 7 time / cyclical features, 6 past utilization metrics and 4 social metrics (e.g. distance to healthcare facilities). Results: XGBoost achieved the best performance with areas under the ROC curve (AUC) of 0.82, 0.81 and 0.78 for 3, 6, 12-month models, and average precision scores (AP) of 0.31, 0.36 and 0.37, respectively, while the commonly used Charlson Comorbidity Index had poor performance with AUCs of 0.63 - 0.64 and APs of 0.1 - 0.17. Past utilization was the best predictor of future utilization. Targeting patients with the top 5 and 10% highest risk for utilization achieved sensitivities of 26 and 40% and positive predictive values of 50 and 38% (12-month model, Figure). Conclusions: Machine learning can be used to predict which patients will have high future healthcare utilization. This may help target populations for intervention programs aimed at reducing utilization.