Machine Learning to Predict Patients at Risk for Frailty in Postoperative Cardiac Patients Using Electrocardiogram and Accelerometer Data

Purpose: Frailty, characterized by diminished physiological reserves, predicts adverse outcomes in older adults undergoing cardiac surgery. This study aimed to identify predictors of frailty using wearable-derived electrocardiogram (ECG) and accelerometry data. The goal was to enhance postoperative risk stratification and rehabilitation. Methods: Data from 80 patients after cardiac surgery were sourced from PhysioNet, including ECG and accelerometry during rehabilitation exercises. The Edmonton Frail Scale dichotomized patients at ≥7 for high frailty. Electrocardiogram features and accelerometry metrics were extracted after signal processing. Missing data were imputed using mean and mode methods, and multicollinearity was addressed. Logistic Regression, Random Forest, and Neural Network models predicted frailty status, with performance evaluated through accuracy, recall, and F1 scores. Results: Frail patients were older (75 vs 72 years, P = .011) and had shorter 6-minute walk distances (259.47 vs 300.92 meters, P = .0424). They also exhibited longer walk times ( P = .007) and higher gait entropy ( P = .0279). Logistic regression achieved the best performance (accuracy = 0.84, recall = 0.59, F1 = 0.61). Key predictors included step length asymmetry, 6-minute walk test time, age, and maximum heart rate (odds ratios >4). The model showed strong discriminative ability (area under the curve = 0.92). Conclusions: Wearable-derived ECG and accelerometry features effectively predict frailty postcardiac surgery. These predictions enable screening and tailored rehabilitation. Future research should explore longitudinal changes and combined feature impacts on clinical outcomes.