Artificial intelligence-enabled electrocardiogram algorithm for simultaneous assessment of left ventricular systolic and diastolic dysfunction

Assessment of left ventricular (LV) systolic and diastolic dysfunction (LVSD and LVDD, respectively) is essential for the evaluation and management of cardiac disease. A rapid, easily performed test to assess both LVSD and LVDD is lacking. We aimed to develop an artificial intelligence-enabled electrocardiogram (AI-ECG) algorithm to simultaneously identify LVSD, defined as LV ejection fraction (EF) <50%, and LVDD, defined as increased left ventricular filling pressure. We also aimed to assess the prognostic performance of 4 groups (normal LV function, LVSD only, LVDD only, and both LVSD and LVDD) determined by a single AI-ECG model classifying both LVSD and LVDD. We trained a multi-output convolutional neural network using 12-lead ECGs performed within 14 days of transthoracic echocardiography from 120,699 patients (98,736 for training and 21,963 for validation) to predict LVSD and LVDD. The trained model was evaluated using the area under the curve (AUC) of the receiver operating curve, accuracy, sensitivity, and specificity. We also assessed whether the model discriminates the risk of all-cause mortality using Kaplan-Meier estimate and Cox proportional hazards regression was used to estimate the hazard. For LVSD, AI-ECG model had the AUC of 0.91, accuracy 93.5%, sensitivity 54.8%, and specificity 97.4%. For LVDD, the model had the AUC 0.91, accuracy 87.3%, sensitivity 59.8%, and specificity 95.1% (Table). When the test patients were separated into 4 groups, the mortality of 4 groups determined from echocardiography (HR 1.8, 95% CI 1.8-1.8) and AI-ECG (HR 1.7, 95% CI 1.7-1.8) was comparable (Figure) with the worst outcome in the group having both LVSD and LVDD. The group with solitary LVDD had worse outcome than solitary LVSD. The AI-ECG simultaneously and rapidly detects LVSD and LVDD with an excellent mortality prognostic information that is equivalent to that found using echocardiography.TablePerformance of AI-ECG for echocardiographically determined LVSD and LVDD.AUCAccuracySensitivitySpecificityLVSD0.9193.5%54.8%97.4%LVDD0.9187.3%59.8%95.1% Open table in a new tab