Web-Based Application of Simplified Machine Learning for Detecting Reduced LVEF From 12-Lead ECG

Deep learning (DL) models have shown high accuracy in detecting reduced left ventricular ejection fraction (LVEF) from electrocardiograms (ECGs). However, their complexity limits clinical use. To address this, we aimed to develop and validate simplified machine learning (ML) models using numerical parameters from 12-lead ECGs to detect LVEF < 40% and to implement them in a user-friendly web application.

We retrospectively analyzed ECG and transthoracic echocardiography data from 21 471 patients across two institutions. The dataset was divided into a development cohort (non-atrial fibrillation [non-AF]: n = 12 922; AF: n = 1281) and an external validation cohort (non-AF: n = 6284; AF: n = 984). Four machine learning algorithms-random forest (RF), extreme gradient boosting (XGBoost), support vector machine, and generalized additive models with LASSO-were evaluated for predicting LVEF as a continuous variable and binary outcome (< 40%).

For continuous LVEF prediction, RF achieved R ² values of 0.68 (non-AF) and 0.74 (AF) in internal validation but performed poorly in external validation. Other models showed R ² values below 0.40 in internal validation. For binary classification, all models achieved area under the curve (AUC) values > 0.90 in the non-AF group during internal validation. RF and XGBoost showed strong performance in the AF group (AUC > 0.90 internally) and adequate accuracy externally (AUCs of 0.80-0.81 in AF and 0.90 in non-AF).

We developed a simple web-based tool for preliminary screening of reduced LVEF using 12-lead ECG parameters.