Cardiac surveillance of childhood cancer using artificial intelligence-enabled electrocardiograms.

Aims: To assess the potential for artificial intelligence-enabled electrocardiogram (AI-ECG) to serve as a long-term cardiac surveillance tool and predict left ventricular systolic dysfunction in childhood cancer patients. Methods and results: We assessed performance of our previously established AI-ECG model to predict left ventricular ejection fraction (LVEF) ≤50% and ≤40% in patients with childhood cancer during internal testing (Boston Children's Hospital) and external validation (Children's Hospital of Philadelphia). The internal test cohort comprised 447 patients [57% male; age at cancer diagnosis 11.2 (5.4-15.7) years; 1553 ECG-echo pairs at median age 13.5 (IQR 7.7-17.9) years; 6.4% with LVEF ≤50%; 1.3% with LVEF ≤40%], 28% with leukaemia, 16% with lymphoma, 8% with neuroblastoma, 8% with sarcoma, 2% with gastrointestinal cancers, 3% with genitourinary cancers, 6% with central nervous system cancers, 11% with other/unspecified cancers, and 18% with missing/unknown cancer labels. Treatment strategies included anthracyclines (35%), bone marrow transplant (7%), and radiation (1%). The external test cohort comprised 2964 patients [55.4% male; 7054 ECG-echo pairs at median age 11.6 (IQR 6.8-15.1) years; 2.5% with LVEF ≤50%; 0.9% with LVEF ≤40%]. Similar AUROCs (0.80-0.85), sensitivities (0.75-0.82), NPVs (0.986-0.996), and percent predicted negative (51-65%) were obtained across institutions to predict LVEF ≤50%, outperforming a biomarker-based model benchmark. Patients with high AI-ECG risk scores for LVEF ≤50% had higher rates of mortality [hazard ratio 3.1 (95% CI 1.8-5.3), P < 0.001] compared to patients with low AI-ECG risk scores. Conclusion: AI-ECG shows promise as a digital biomarker for cardiac surveillance in the vulnerable childhood cancer survivor population.