Artificial intelligence–enhanced electrocardiography for detection of subclinical left ventricular dysfunction in cancer survivors
TL;DR
Artificial intelligence-enhanced electrocardiogram shows moderate discrimination for subclinical HF, with significant improvement in predictive accuracy when combined with CHERISH risk score among cancer survivors, and compares it with clinically-derived and electronic medical record-based HF risk scores.
Abstract
Cancer survivors face elevated heart failure (HF) risk, with subclinical HF often detectable by reduced global longitudinal strain (GLS). Strain imaging is resource-intensive and not universally accessible. Artificial intelligence-enhanced electrocardiogram (AI-ECG) offers a scalable alternative, but its performance for detecting reduced GLS in survivors remains unclear. To evaluate the discriminatory performance of AI-ECG for detecting reduced GLS in asymptomatic cancer survivors and compare it with clinically-derived and electronic medical record (EMR)-based HF risk scores. Asymptomatic cancer survivors ≥5 years post-treatment underwent paired ECG and echocardiography with GLS assessment. Reduced GLS was defined as >-16%. Nine machine learning models were assessed using 70% of data for training and cross-validation and 30% for independent testing. From 1,900 ECG-derived waveform features, 50 ANOVA-selected features yielded optimal discrimination. ARIC-HF (incorporating clinical examination and blood biomarkers) and CHERISH (an EMR-based score using routinely coded diagnoses) risk scores were calculated using published coefficients. Discrimination was assessed by area under the receiver operating characteristic curve (AUC). AI-ECG was evaluated alone, combined with each score via percentile rank averaging, and within score-defined risk strata. A total of 472 cancer survivors (mean age 64.8 years; 86% female) were included; of which 80 (16.9%) had reduced GLS. AI-ECG achieved the highest discrimination (AUC 0.791; sensitivity 0.958, specificity 0.585; negative predictive value 0.986, positive predictive value 0.319), outperforming ARIC-HF (AUC 0.713) and CHERISH (AUC 0.671) in the testing dataset. Combining AI-ECG with the CHERISH risk score improved discrimination (AUC 0.806; ΔAUC +0.14, 95% CI 0.06–0.21), with the strongest performance observed in low-risk strata (AUC 0.863). ARIC-HF combined with AI-ECG showed no significant improvement in predictive performance (AUC 0.793). AI-ECG shows moderate discrimination for subclinical HF, with significant improvement in predictive accuracy when combined with CHERISH risk score among cancer survivors.Triage Performance for Reduced GLS DetectionDiagnostic Metrics for Reduced GLS Detection