Identifying Ventricular Dysfunction Indicators in Electrocardiograms via Artificial Intelligence-Driven Analysis

Recent studies highlight artificial intelligence's ability to identify ventricular dysfunction via electrocardiograms (ECGs); however, specific indicative waveforms remain unclear. This study analysed ECG and echocardiography data from 17,422 cases in Japan and Germany. We developed 10-layer convolutional neural networks to detect left ventricular ejection fractions below 50%, using four-fold cross-validation. Model performance, evaluated among different ECG configurations (3 s strips, single-beat, and two-beat overlay) and segments (PQRST, QRST, P, QRS, and PQRS), showed two-beat ECGs performed best, followed by single-beat models, surpassing 3 s models in both internal and external validations. Single-beat models revealed limb leads, particularly I and aVR, as most indicative of dysfunction. An analysis indicated segments from QRS to T-wave were most revealing, with P segments enhancing model performance. This study confirmed that dual-beat ECGs enabled the most precise ventricular function classification, and segments from the P- to T-wave in ECGs were more effective for assessing ventricular dysfunction, with leads I and aVR offering higher diagnostic utility.