China-calibration-ready echocardiographic feature modeling for LVEF-derived dysfunction classification: a surrogate-label proof-of-concept

Abstract

Introduction This study presents a reproducible echocardiographic feature-engineering and machine-learning pipeline for classifying a left ventricular ejection fraction (LVEF)-derived surrogate dysfunction label, using publicly available two-dimensional echocardiographic tracings. The work is explicitly framed as a methodological proof-of-concept rather than a clinically validated model for predicting anthracycline-related right ventricular cardiotoxicity. Methods A derivative dataset comprising 10,930 records was assembled from the EchoNet-Dynamic database and a CAMUS-derived frame-mask archive. From this pool, a stratified 1,000-record working sample was extracted to support auditable and transparent model development. After mutual-information ranking performed within the training folds, twelve geometric and morphometric features were retained, including left ventricular trace-area change, bounding-box area change, and cohort-median-normalized calibration variables-these were deliberately designed for later substitution with age- and sex-specific Chinese reference values from the EMINCA-II study. Five classifiers were trained using an 800/200 stratified split, with five-fold cross-validation applied inside the training set. Results Among the five classifiers, logistic regression achieved the highest held-out ROC-AUC of 0.996 and F1 score of 0.966. However, these performance metrics reflect the geometric coupling between the engineered left-ventricular features and the LVEF-derived label, rather than independent clinical predictive capability. A descriptive mediation analysis further showed that the association between bounding-box area change and the dysfunction label was largely transmitted through left ventricular trace-area change. Discussion The available public data do not include information on anthracycline exposure, Chinese patient identity, right ventricular functional outcomes, or prospective cardiotoxicity endpoints. Therefore, the primary contribution of this study is a transparent, calibration-ready analytical scaffold that can be reused and extended in future prospective Chinese anthracycline cohorts that incorporate true right ventricular outcome measures.