What question did this study set out to answer?

The aim was to develop machine-learning models using ECG features to predict stroke risk over the next 15 years.

May 20, 2026

A80-1-05 Predicting Long-term Stroke Occurrence Using Ecg Signals From a Single-night Polysomnography

SSS SahaMeharry Medical College EBE BaileyMeharry Medical College

Key Points

The aim was to develop machine-learning models using ECG features to predict stroke risk over the next 15 years.
Utilized ECG recordings and polysomnographic data from the Sleep Heart Health Study with 637 participants.
Generated CNN-based ECG features from four respiratory phases and combined with handcrafted features for analysis.
Applied ten classifiers to the data using GridSearchCV and 3-fold StratifiedKFold cross-validation.
Random Forest model achieved the highest F1 score of 63.6% and recall of 77.8%.
KNN had the highest accuracy at 63.9%, while CatBoost showed comparable recall but lower precision at 51.9%.
ECG features effectively differentiated stroke participants from controls, enhancing cerebrovascular risk screening.

Abstract

Abstract Rationale Sleep-disordered breathing (SDB) leads to intermittent hypoxia and autonomic fluctuations that alter cardiac electrophysiology, leaving measurable imprints on electrocardiographic (ECG) signals. Subtle ECG changes during apneic and hypopneic events may reflect early cerebrovascular vulnerability. This study aimed to develop machine-learning models using ECG features from sleep studies to predict the occurrence of stroke within the next 15 years of the baseline assessment. Methods This retrospective analysis utilized ECG recordings and polysomnographic annotations from the Sleep Heart Health Study (SHHS). A total of 637 participants were analyzed (stroke n = 300, propensity matched controls n = 337; mean age = 73 ± 8 years; BMI = 27.5 ± 4.5 kg/m²; mean AHI = 18.2 ± 10.6 events/hour; 59 % male; 89 % White).ECG features were derived across four respiratory phases: Normal breathing, Entering Apnea/Hypopnea Events, During Events, and Exiting Events, and then averaged at the subject level. For each state, ECG segments were transformed into CNN-based representations (F0-F16) capturing temporal-spectral morphology. Additional handcrafted features, including wavelet-energy coefficients, spectral entropy, heart rate variability, and power ratios, were combined. All features were standardized, non-informative identifiers removed, and highly correlated variables (r 0.9) excluded to prevent leakage.Ten classifiers were trained using GridSearchCV with 3-fold StratifiedKFold cross-validation: Logistic Regression, Decision Tree, Random Forest, Naïve Bayes, K-Nearest Neighbors, Support Vector Machines (linear and RBF), LightGBM, XGBoost, and CatBoost.Models were evaluated on a held-out test set (18 stroke, 18 control) using accuracy, F1-score, recall, precision, and ROC AUC. Results Random Forest produced the highest test F1 = 63.6 %, recall = 77.8 %, and balanced accuracy = 55.6 %. CatBoost achieved comparable recall (77.8 %) but lower precision (51.9 %), while KNN had the highest accuracy (63.9 %). Conclusions We showed that CNN-derived ECG features across apnea-hypopnea transitions can differentiate stroke from control participants. Integrating AHI and ECG biomarkers may enable non-invasive, scalable screening for cerebrovascular risk in sleep-medicine populations. This abstract is funded by: NIH AIM AHEAD, Grant No-1OT2OD032581-02-838

Ask AI

Helpful

Bookmark

View Full Paper