Prediction of left ventricular systolic dysfunction from chest radiographs using validated abnormality labels plus machine learning versus end-to-end deep learning

Although recent deep learning approaches show promise for extracting predictive features from chest radiography (CXR), existing end-to-end models are often difficult to interpret clinically, and real-world prospective validation remains limited. We explored a hybrid deep learning strategy to predict reduced left ventricular ejection fraction (LVEF) from CXR data. We analyzed 14,314 chest radiographs and transthoracic echocardiography (TTE) reports across a derivation cohort (n=10,152), a prospective validation cohort (n=505), and a retrospective external validation cohort (n=3,657). We applied a deep-learning algorithm (Lunit INSIGHT CXR) to detect radiological abnormalities. Subsequently, we trained chest radiography (CXR)-based LVEF prediction models (LVEF ≤ 40% and ≤ 30%) using eXtreme Gradient Boosting (XGBoost) machine learning (hybrid deep learning; HDL) and direct deep learning neural network (DDL). Model performance was assessed using area under the receiver-operating characteristic (AUROC) and precision-recall curves (AUPRC). Local Interpretable Model-Agnostic Explanations (LIME) method was applied to improve XGBoost model interpretability. The HDL approach using validated CXR abnormality features achieved an AUROC of 0.85 and AUPRC of 0.34 in the internal test set, with higher discrimination than DDL (AUROC: 0.61; AUPRC: 0.09). The HDL model also performed well in the prospective validation cohort (AUROC: 0.85). External validation, using the HDL approach only, maintained discriminative performance (AUROC: 0.73). LIME and SHAP analyses identified cardiomegaly, pleural effusion, and pulmonary consolidation as dominant predictive features. This CXR-based model using validated feature extraction showed higher discrimination than the direct deep learning model evaluated in this study and may serve as an interpretable screening aid for cardiac function assessment. Clinical trial registration: NCT04996381.