ECG Signatures and Long-Term Ischemic Stroke Risk: A Deep Learning Analysis of 200,000 Patients

Study Population

ECG2Stroke was developed using data from 2 previously described patient cohorts, the Community Care Cohort Project (C3PO)¹⁶ and the Enterprise Warehouse of Cardiology (EWOC)¹⁷, representing approximately 520,000 patients receiving longitudinal primary care and 99,000 patients receiving longitudinal cardiology care, respectively, within the Mass General Brigham (MGB) multi-institutional healthcare system in the New England region of the US. Briefly, patients aged 18 to 90 years receiving longitudinal primary or ambulatory cardiology care within one of seven MGB network hospitals sharing a unified electronic health record (EHR) between 2000 and 2019 were included. As described in their original publications, patients needed to have at least 2 qualifying visits between 1–3 years apart within primary care (C3PO) or ambulatory cardiology (EWOC) clinics for inclusion, resulting in a longitudinal data structure previously shown to demonstrate favorable data density with less selection bias compared to traditional EHR sampling methods.

The deep learning neural network was trained on patients with ≥1 ECG performed at Massachusetts General Hospital (MGH) within 3 years prior to the start of cohort follow-up (defined as the second of the first qualifying pair of visits required for C3PO or EWOC inclusion). A validation set of patients from MGH without prior history of stroke was used for hyperparameter tuning and model selection. A neural network-derived model was then tested on independent test sets of stroke-free individuals with eligible ECGs from MGH (internal test set) and Brigham and Women’s Hospital (BWH, external test set). Patients with ECGs from both hospitals (4,600 individuals) retained only their MGH-performed ECGs and were only included in the MGH cohort, without inclusion in the BWH cohort. There was therefore no overlap at the individual or ECG level between the training, validation, or test sets. We performed additional external validation among patients receiving longitudinal primary care at Beth Israel Deaconess Medical Center, an independent healthcare system, and with no history of stroke (BIDMC, external test set) (Figure 1).

Although all located within the larger Boston region, MGH, BWH, and BIDMC have distinct patterns of community-based and outside referrals, case mix, and health care utilization as outlined in the 2022 Massachusetts Center for Health Information and Analysis report¹⁸. BIDMC is further distinct as part of an entirely separate health system. For example, top community discharges in 2020 were as follows: BWH (Jamaica Plain, Brookline, Roslindale); MGH (Charlestown, Chelsea, Revere); BIDMC (Brookline, Jamaica Plain, Roxbury). Similarly, top discharge diagnoses by diagnosis-related group were as follows: BWH (craniotomy, cardiovascular procedures without acute myocardial infarction, major small and large bowel procedures), MGH (moderately extensive procedure unrelated to principal diagnosis, craniotomy, major respiratory infection & inflammation, percutaneous cardiovascular procedures without acute myocardial infarction); BIDMC (normal neonate birth, vaginal delivery, cesarean, and cardiovascular procedures without acute myocardial infarction).

ECG Acquisition and Neural Network Training

The input to the deep learning neural network is a single 12-lead ECG, representing a 10-second time-series sampled at 500 Hz for each of the 12 leads. To create a uniform input tensor of dimension 5000 × 12, ECGs with lower sampling rates (a minority) were linearly up-sampled to 500 Hz, and those with shorter durations were zero-padded to 5000 measurements. Raw waveform data and tabular metadata (including date, time, machine type, sampling frequency, and automated and physician reads) were extracted from all ECGs. All ECGs had been performed for clinical purposes. No ECGs were excluded from the training or test sets on the basis of particular clinical findings.

The neural network used a loss function that accounted for both time-to-outcome (i.e., ischemic stroke) and missingness due to censoring (defined as the earliest of death or last encounter). This loss function was previously used to predict 5-year AF-free survival¹². Similar to prior work, we also trained the network to perform related secondary tasks (prediction of 10-year overall survival, age estimation, sex classification, and identification of AF on the ECG). This multitask approach improved stroke prediction performance compared with other model-building approaches considered during model derivation (Table S1). Further details are supplied in the Supplemental Methods.

The neural network was trained using all 12-lead ECGs in the MGH training dataset performed within 3 years before the start of follow-up. Since ECGs from individuals with prevalent disease are likely to contain rich training signals, patients with prior stroke were included in the training set, consistent with prior studies¹². However, individuals with prevalent stroke were excluded from all validation and test sets.

The neural network is a deep, 1-dimensional convolutional network. The architecture is summarized in Figure S1. Briefly, the network consists of multiple densely connected convolutional blocks followed by several fully-connected layers. These layers combine information from the entire 12-lead ECG to produce a longitudinal estimate of stroke-free survival. Learning curves are shown in Figure S2, and the performance of the neural network on secondary tasks is shown in Figure S3.

Clinical Factors

Baseline age, sex, and race were obtained from the EHR. Race/Ethnicity was classified as White, Black, Hispanic/Latino, Asian or Pacific Islander, Other/Unknown. Tobacco use was categorized as present or absent. The presence of hypertension, hyperlipidemia, diabetes, chronic kidney disease, heart failure, valvular disease, and myocardial infarction were ascertained using previously published sets of diagnostic codes and medications (eg, antiglycemics), having a reported positive predictive value (PPV) ≥85% in MGH and BWH^19,20. AF was defined using a previously published algorithm comprising diagnostic and procedural codes and ECG reports (PPV 92%)²¹. Antihypertensive use was determined using medication lists¹⁹. The presence of obesity was similarly defined using diagnosis codes. Use of oral anticoagulants and statin drugs was determined using medication lists (for oral anticoagulants, outpatient only). Blood pressure values, LDL, lab values were obtained from the EHR. Analogous definitions were applied in BIDMC. Disease definitions are listed in detail in Table S2.

Outcomes

The primary outcome was incident ischemic stroke, defined using previously published diagnosis code sets having PPV ≥85% in MGH and BWH and strong pairwise agreement between manual reviewers (Cohen’s κ = 0.8), modified to remove codes specific for transient ischemic attack (TIA)¹⁹. The entirety of the EHR (regardless of treating site) was used for outcome ascertainment. A similar definition was employed in BIDMC (Table S2).

Statistical Analysis

Our primary objective was prediction of incident ischemic stroke at 10 years in the MGH, BWH, and BI test sets. Shorter time horizons at 1 and 5 years were also assessed. During model evaluation, patients with stroke prevalent at the start of follow-up were excluded and only the most recent ECG before start of follow-up was used for each individual. To avoid misclassification of prevalent strokes as incident due to delayed coding, a 30-day blanking period was applied such that person-time began 30 days after the start of cohort follow-up. Events occurring during the 30-day blanking period were considered prevalent. To correct for miscalibration introduced by inclusion of patients with prevalent stroke in the neural network training process, we fit Cox proportional hazards models including neural network-predicted event probabilities alone or with age and sex as covariates for 1, 5, and 10-year horizons in the subset of the combined training and validation sets (MGH development set) with prevalent stroke excluded. Predicted event probabilities were logit-transformed to achieve an approximately linear relationship with the log hazard²². Given a low death rate when considering person-time prior to stroke or within 365 days of the last encounter (8.8% MGH development set), death was treated as a censoring event. Sensitivity analyses were subsequently performed using Fine-Gray models and the cumulative incidence function with death as a competing risk. Given our intent to assess the feasibility of population-level stroke risk stratification using a single 12-lead ECG, and the universal availability of age and sex in this context, the Cox model combining neural network-predicted stroke probabilities with age and sex was selected as our primary model, hereafter referred to as ECG2Stroke. A statistically significant interaction was observed during training between neural network predicted probabilities and male sex (hazard ratio [HR] 0.91, [95% CI, 0.85–0.98]), but the interaction was not included in the final model given marginal improvement in Akaike Information Criterion. There was no significant interaction with age. ECG2Stroke model parameters are provided in Table S3A. Although used as a secondary training target for the neural network, presence of AF in the diagnostic statement of the qualifying ECG as an additional covariate was not a significant predictor and caused minimal change in remaining ECG2Stroke regression coefficients, and therefore was not included in the final model.

To compare the performance of ECG2Stroke to the revised FSRP⁷, we further restricted the MGH development set to patients with complete data (22%) for calculation of the FSRP and fit analogous Cox proportional hazard models, but including as covariates predicted event probabilities from ECG2Stroke alone, FSRP alone, and both ECG2Stroke and FSRP. Additional baseline models based on ECG alone or age and sex were also fit in the same population. No significant interaction was observed between neural network probabilities with either age, sex, or FSRP. Model fit parameters for ECG2Stroke in the FSRP complete data subset are provided in Table S3B.

Discrimination and calibration were evaluated in the test sets by calculating the time-dependent area under receiver operating characteristic curve (AUROC)²³, time-dependent average precision (AP)²⁴, and integrated calibration index²⁵ (ICI) fit using restricted cubic splines. AUROC and AP estimates were calculated using inverse probability of censoring weights. AP was normalized to the baseline event rate (AP of a non-informative model) and reported as nAP. P-values for AUROC and AP comparisons and confidence intervals were obtained by bootstrap. The same evaluation metrics were assessed in subgroups of age, sex, race, and clinical risk factors in external test sets (BWH and BIDMC test sets). Incremental model contribution was quantified using variance decomposition in Cox models, reporting partial R² derived from likelihood-ratio testing to estimate the proportion of explained variation²⁶. We explored stroke-free survival in high and low risk strata predicted by ECG2Stroke vs FSRP models using risk thresholds (low risk: predicted 10-year risk <5%; high risk: >20%) derived from standard atherosclerotic cardiovascular disease (ASCVD) prevention guidelines²⁷. Survival differences were compared using log-rank testing. Reclassification at 10-year horizon was quantified with time-dependent net reclassification improvement (NRI) using the inverse probability weighting method²⁸ and confidence intervals obtained by bootstrap. Decision curves were constructed across a range (0.01–0.45) of predicted 10-year risk thresholds.²⁹ We evaluated the decision curves at selected thresholds of 10-year risk (0.05,0.075,0.1 0.2) to report net benefit and net interventions avoided and quantified the corresponding classification performance (sensitivity, specificity, positive predictive value, and negative predictive value) implied by the decision rules. We computed the linear predictors of ECG2Stroke and FSRP-derived Cox in the test sets and assessed Pearson’s correlation.

Secondary analyses and saliency

Additional analyses examined the performance of ECG2Stroke across strata of prevalent AF (MGH, BWH, and BIDMC test sets) and baseline oral anticoagulation (OAC) use (MGH and BWH only). Among patients with prevalent AF, stroke-free survival and reclassification was examined for high- and low-risk strata defined using guideline-based thresholds according to the CHA₂DS₂-VASc³⁰ score (Low risk: <1 for males and <2 for females, High risk: >1 for males and >2 for females) versus high- and low-risk strata defined using the ECG2Stroke thresholds listed above. To ensure adequate sample size, the MGH and BWH test sets were combined into the MGH/BWH test set for analyses restricted to specific subgroups such as the FSRP complete data subset or prevalent AF subset. Decision curves for CHA₂DS₂-VASc mapped integer scores to corresponding Kaplan-Meier estimates. We further modeled AF status as a time-dependent variable alone or jointly with ECG2Stroke to assess marginal contributions to risk prediction by inspecting their interaction, coefficient attenuation, and partial R².

To assess electrocardiographic features driving the neural network’s predictions, we plotted median ECG waveforms and produced saliency maps using a random sample of 7,080 individuals from the BWH test set. These saliency maps depict areas of the ECG where changes in voltage had the greatest influence on the model’s predicted risk estimates. Further details of saliency map and median waveform analyses are described in the Supplemental Methods. To further characterize the neural network, we calculated Pearson correlation of its output probabilities (logit transformed) to known P-wave indices (e.g., PR interval, P-wave duration).

We assessed ECG2Stroke and FSRP outcome specificity by examining discrimination (AUROC) and calibration (ICI) performance for incident heart failure (HF), myocardial infarction (MI), and death across all test sets. In a small subset of patients with incident strokes with manual physician-adjudicated stroke subtype information, cause-specific Cox models were used to estimate associations with cardioembolic vs non-cardioembolic stroke subtypes, with heterogeneity assessed using the Lunn-McNeil method³¹. Further details of the stroke subtyping labels are provided in the Supplemental Methods. We characterized ECG2Stroke risk patterns by assessing associations between its output probabilities (logit transformed) and the presence or absence of baseline clinical stroke risk factors using age- and sex-adjusted linear regression models. Analyses were performed using Python v3.13³² and R v4.5³³.

Data Sharing

The ECG2Stroke AI model and weights are available under an academic use license at https://github.com/broadinstitute/ml4h/tree/master/model_zoo/ECG2Stroke.

Training and Validation Samples

A total of 107,513 patients from MGH were further subdivided into training (n=91,306) and validation (n=10,190) sets used for model development, as well as an internal test (n=4,771) set. The training set included 299,640 12-lead ECGs (median 6 ECGs per individual). The MGH training and validation sets were collapsed into a single MGH development set (n=101,496) for parameterization of Cox models. Separate external test sets were constructed from patients with ECGs performed at BWH (n=68,884) and BIDMC (n=29,882). Figure 1 depicts a study flow diagram. The proportion of patients with an incident stroke event within 10 years was 4.5% in the MGH development set over median follow-up duration 7.0 years, 7.3% over 7.5 years in the MGH internal test set, 4.7% over 7.0 years in the BWH external test set, and 4.1% over 10.8 years in the BIDMC external test set. Baseline sample characteristics are summarized in Table 1.

Model Performance in Test Sets

ECG2Stroke demonstrated strong univariable association with 10-year incident stroke in the test sets (HR per SD: MGH, 2.838 [95% CI 2.482–3.245]; BWH, 2.648 [2.535–2.765]; BIDMC, 2.803 [2.609–3.012]) and exhibited moderate discrimination of incident stroke (10-year AUROC: MGH, 0.795 [0.768–0.820]; BWH, 0.774 [0.765–0.783]; BIDMC, 0.772 [0.758–0.785]) (10-year nAP: MGH, 1.970 [1.720–2.280]; BWH, 1.760 [1.680–1.860]; BIDMC, 1.950 [1.810–2.120]) (Table 2). Estimates were generally well-calibrated (10-year horizon ICI: MGH, 0.030; BWH, 0.005; BIDMC, 0.026; Table 2, Central Illustration), and stratified longitudinal stroke risk (Central Illustration). AUROC values were largely similar at shorter time horizons (5-year horizon range: 0.753–0.765; 1-year horizon: 0.701–0.749; Tables S4A,B), while nAP improved with shorter time horizons (10-year horizon range: 1.760–1.970; 5-year horizon: 2.110–2.450; 1-year horizon: 2.470–3.190; Table 2, Tables S4A,B).

ECG2Stroke demonstrated consistently higher average precision compared to a model based on age and sex alone and generally favorable AUROC compared to models based on ECG alone or age and sex alone (Table 2, Table S4A). Variance decomposition confirmed the ECG alone model explained a significant proportion of additional variance in 10-year stroke risk beyond age and sex alone (Partial R²: ECG only, 0.082–0.118, Table S5). In both BWH and BIDMC, ECG2Stroke stratified risk across subgroups of sex, age, race, and clinical risk factors while maintaining discrimination and calibration performance (Figure 2, Figures S4–S5, Tables S6A,B). Decision curves for ECG2Stroke demonstrated net benefit across a range of clinically relevant thresholds (Figure S6). Sensitivity, specificity, positive and negative predictive values for a range of clinically relevant thresholds are included in Table S7. Results were similar when death was modeled as a competing risk, where specifically we observed modest attenuation of effect size (10-year horizon ECG neural network hazard ratio: death as censoring, 2.52 [2.44–2.60]; death as competing risk, 2.42 [2.34–2.49]), while absolute 10-year stroke risk estimates using the Aalen–Johansen estimator (AJ) were essentially identical to Kaplan–Meier (KM) estimates (MGH test: KM, 0.104, AJ, 0.103; BWH test: KM, 0.072, AJ, 0.071; BIDMC test: KM 0.053, AJ, 0.052).

Comparison to FSRP

We then compared the performance of ECG2Stroke to the FSRP alone and their combination, among individuals with available data. In fitting the model combining both ECG2Stroke and FSRP, both the neural network-predicted probabilities and FSRP were independently associated with incident stroke (neural network probabilities: HR, 2.06 [1.84–2.30]; FSRP: HR, 1.62 [1.37–1.93]), without significant interaction.

To ensure sufficient sample size, assessment of FSRP was performed in a combined MGH/BWH test set along with the BIDMC test set (baseline characteristics in Table S8). Here, both ECG2Stroke and FSRP demonstrated strong univariable association with 10-year incident stroke in the test sets (MGH/BWH test HR per SD: ECG2Stroke, 3.016 [2.795–3.253]; FSRP, 2.732 [2.534–2.945]; BIDMC test HR per SD: ECG2Stroke, 2.474 [2.269–2.698]; FSRP, 2.301 [2.118–2.500]; Table 3). The discrimination and calibration of ECG2Stroke at a 10-year horizon was comparable to the FSRP-based model but with numerically higher point estimates (MGH/BWH test AUROC: ECG2Stroke, 0.791 [0.776–0.806]; FSRP, 0.779 [0.763–0.795]; BIDMC test AUROC: ECG2Stroke, 0.745 [0.727–0.763]; FSRP,0.728 [0.710–0.747]). Discrimination was marginally improved in the model combining ECG2Stroke and FSRP (10-year AUROC: MGH/BWH test, 0.800 [0.785–0.815]; BIDMC test, 0.754 [0.735–0.772]; Table 3). Calibration was favorable in the MGH/BWH test set compared to the BIDMC test set, with similar behavior using ECG2Stroke and FSRP (Table 3, Figure 3). Similar patterns were observed at shorter time horizons (Tables S9A,B). Both models retained predictive performance across subgroups of age, sex, race, and clinical risk factors in both test sets (Figure 2, Tables S10A,B). Sensitivity analysis restricting the Cox model training and testing cohorts to age >55 years to match the original FSRP validation cohort resulted in worse FSRP 10-year performance across test sets (AUROC: MGH/BWH test, 0.737 [0.714–0.757]; BIDMC test, 0.681 [0.660–0.703]; nAP: MGH/BWH test: 1.420 [1.340–1.530]; BIDMC test, 1.590 [1.440–1.810]; ICI: MGH/BWH test: 0.013; BIDMC test, 0.050).

Using low (10-year risk <5%) and high (10-year risk >20%) risk thresholds derived from standard ASCVD prevention guidelines²³, we found similar stratification of longitudinal stroke incidence among ECG2Stroke and FSRP-defined risk groups (MGH/BWH test: Low risk, p = 0.82; High risk, p = 0.86; BIDMC test: Low risk, p = 0.23; High risk, p=0.97 Figure 3). ECG2Stroke improved reclassification modestly at the high vs intermediate risk threshold (MGH/BWH test: NRI overall, 0.048 [95% CI, 0.020–0.075]; (NRI+), 0.054 [0.027–0.081]; (NRI−), −0.006 [−0.010–0.002]; BIDMC test: NRI overall, 0.073 [0.044–0.104]; (NRI+), 0.092 [0.063–0.123], (NRI−), −0.019 [−0.024 – −0.015]). Reclassification was not significantly improved at the low vs intermediate risk threshold (MGH/BWH test: NRI overall, 0.002 [−0.029–0.030]; (NRI+), 0.002 [−0.026–0.028]; (NRI−), 0.000 [−0.012–0.012]; BIDMC test: NRI overall, 0.016 [−0.010–0.041]; (NRI+), 0.034 [0.010–0.057], (NRI−), −0.019 [−0.026 - −0.011]). Findings were similar using alternate risk thresholds (low: 10-year risk <7.5%; high: >10%) (Figure S7). Decision curves for 10-year stroke demonstrated modest improvements in net benefit using ECG2Stroke compared to FSRP across a range of relevant thresholds (threshold probabilities: MGH/BWH test, 1%–40%; BIDMC test, 1%–14%) (Figure S8). Sensitivity, specificity, positive and negative predictive values along with net benefit values for a range of clinically relevant thresholds for both models are included in Table S11.

The linear predictors of ECG2Stroke and FSRP-based model were highly correlated in all test sets (Pearson correlation: MGH/BWH test set, 0.82; BIDMC, 0.81; p<0.01 for both). However, variance decomposition confirmed both ECG2Stroke and FSRP explain a significant proportion of additional variance in 10-year stroke risk beyond the other alone (MGH/BWH Partial R²: ECG2Stroke, 0.076; FSRP, 0.016; BIDMC Partial R²: ECG2Stroke, 0.057; FSRP, 0.024) (Table S12).

Atrial Fibrillation

We examined the effect of AF on model performance given its particular relevance to both ischemic stroke risk and the ECG. In the combined MGH/BWH test set, ECG2Stroke discrimination and calibration were more favorable among individuals with versus without prevalent AF (Table 4, Tables S13A,B, Figure S9). Calibration was again favorable in the MGH/BWH test set compared to the BIDMC test set. Stratification by OAC status in MGH/BWH revealed no major changes in model performance among patients with AF. Similar trends were observed at shorter time horizons. To assess the extent to which ECG2Stroke performance extends beyond prediction of AF, we examined model performance among patients without AF at baseline with censoring at incident AF diagnosis, observing continued moderate discrimination and good calibration (Table 4, Tables S13A,B).

In an exploratory analysis, we compared classification of AF patients not on OAC at baseline into low and high stroke risk groups using standard risk thresholds for CHA₂DS₂-VASc scores (low: <1 for males and <2 for females, intermediate: 1 for males and 2 for females, high: >1 for males and >2 for females) and using standard ASCVD prevention risk thresholds (low: 10-year risk <5%, high: >20%) for ECG2Stroke in the MGH/BWH test set (Figure 4). Stroke-free survival differed between the high-risk groups (log rank p<0.005), but not the low-risk groups (p=0.43). The ECG2Stroke low-risk group included more patients (n=1096) than the corresponding CHA₂DS₂-VASc group (n=288), while the reverse was true for high-risk groups (ECG2Stroke n=1804, CHA₂DS₂-VASc n=4981). Quantiles of ECG2Stroke were also able to stratify risk further within CHA₂DS₂-VASc-defined high-risk and intermediate-risk groups (Figure 4). Results were similar if patients were censored at OAC initiation (Figure S10) or with use of alternate risk thresholds (Figure S11). ECG2Stroke applied at standard risk thresholds improved reclassification at the low versus intermediate risk threshold (NRI overall, 0.119 [0.082–0.155]; (NRI+), −0.068 [−0.098- −0.043]; (NRI−), 0.187 [0.165–0.210]) and at the high versus intermediate threshold (NRI overall, 0.071 [0.018–0.126]; (NRI+), −0.546 [−0.590- −0.495]; (NRI−), 0.616 [0.593–0.640]), driven primarily by appropriate downclassification of non-events. Use of the less stringent alternate risk thresholds (low: 10-year risk <7.5%; high: >10%) for ECG2Stroke further increased reclassification improvement (low versus intermediate threshold NRI: overall, 0.206 [0.161–0.248]; (NRI+), −0.141 [−0.177- −0.107]; (NRI−), 0.347 [0.322–0.372]; high versus intermediate threshold NRI: overall, 0.117 [0.069–0.166]; (NRI+), −0.162 [−0.201- −0.121]; (NRI−), 0.279 [0.252–0.304]). Decision curves for 10-year stroke demonstrated substantial improvements in net benefit using ECG2Stroke compared to CHA₂DS₂-VASc across a range of clinically relevant thresholds (threshold probabilities: MGH/BWH test, 1%−24%) (Figure S12). Sensitivity, specificity, positive and negative predictive values along with net benefit values for a range of clinically relevant thresholds for both models are included in Table S14.

Time-dependent modeling of AF status to directly characterize relations between ECG2Stroke, AF, and 10-year stroke risk revealed greater coefficient attenuation with joint modeling for AF than ECG2Stroke, much larger partial R² for ECG2Stroke, and a moderate submultiplicative interaction between the two (HR for interaction range: 0.658 – 0.745) (Figure 5, Tables S15A,B). Results were similar when we accounted for potential occult subclinical AF by treating AF detected up to 2 years post-stroke as prevalent AF (Tables S15A,B).

Clinical and electrocardiographic correlates of ECG-based stroke risk

Examination of ECG neural network saliency maps demonstrated that variation in the ECG waveform near the region of the P-wave had the greatest effect on model predictions (Central Illustration). We examined associations of the ECG neural network component of ECG2Stroke with available ECG-based atrial substrate markers and found moderate correlation with PR interval (Pearson’s R: MGH development, 0.300 MGH test, 0.281; BWH test, 0.298; p<0.001 for all) (Figure 6) and modest but significant correlations with P-wave duration (Pearson’s R: MGH test, 0.039; BWH test, 0.033; p<0.001 for all) and P-wave axis (Pearson’s R: MGH test, 0.034; BWH test, 0.044; p<0.001 for all)(Figure S13). A Cox model including all 3 P-wave ECG metrics (PR interval, P-wave duration, P-axis) as predictors trained in the MGH development set retained some discriminative performance for stroke in the test sets but was significantly worse than models based on ECG or Age+Sex (Table S16).

To quantify model specificity, we examined associations between stroke risk models and alternative incident cardiovascular outcomes including HF, MI, and death. Here, although both ECG2Stroke and FSRP discriminated each of the alternative outcomes (Table S17), there was substantially better agreement between observed and expected probabilities (i.e., calibration) for incident stroke when compared with HF and death, although generally comparable for MI (Figure S14). We further quantified associations with cardioembolic (CE) vs non-cardioembolic (non-CE) stroke in a small subset of patients with incident strokes which were manually subtyped (n CE=119, n non-CE=139). ECG2Stroke-predicted 10-year stroke risk was strongly associated with CE stroke but not non-CE stroke (cause-specific HR per 1-SD increase in linear predictor: CE, 2.168 [1.641 – 2.865]; non-CE, 1.133 [0.923 – 1.391]; p-value for heterogeneity <0.001) (Figure 7). Consistent with this, patients with ECG2Stroke-predicted 10-year stroke risk below the median had disproportionately more non-CE strokes (bottom 50% predicted stroke risk: CE 71/129 [55%], top 50% predicted stroke risk: CE: 48/129 [37%]; χ² 7.549, p<0.01). Among baseline clinical factors, ECG2Stroke demonstrated strongest associations with HF, AF, and chronic kidney disease (Figure 8).