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; 1,553 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 leukemia, 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 2,964 patients (55.4% male; 7,054 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 (HR 3.1 [95% 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.