An ECG-based interpretable machine learning approach for the differential diagnosis of cardiac amyloidosis and hypertrophic cardiomyopathy

Background Differentiating cardiac amyloidosis (CA) from hypertrophic cardiomyopathy (HCM) in patients with ventricular hypertrophy is clinically crucial but often challenging. Advanced imaging techniques provide unique diagnostic information but are not universally available. Electrocardiography (ECG), a low-cost and widely accessible tool, may offer valuable information to guide early diagnostic prioritisation. Purpose To develop and validate an interpretable, ECG-based machine-learning approach to differentiate CA from HCM using automatically extracted morphological ECG features. Methods We retrospectively analysed digital 12-lead ECGs from 161 patients with ventricular hypertrophy (25 CA, 136 HCM). A total of 468 ECG morphological variables were automatically extracted using the Philips DXL algorithm. After preprocessing, dimensionality reduction, and correlation filtering, feature selection was performed using a gradient boosting machine (XGBoost) across 300 resampling iterations with class balancing. Selected features were incorporated into an interpretable logistic regression–based risk score. Model performance was assessed using Harrell’s bootstrap validation and compared with the classical low-voltage ECG criterion. Results Three ECG parameters consistently emerged as key discriminators: R-wave amplitude in lead I, R-wave amplitude in lead II, and ventricular activation time in lead II (table 1). The resulting ECG-based score demonstrated excellent discrimination between CA and HCM (bootstrap-corrected AUC 0.94; Brier score 0.055), significantly outperforming the classical low-voltage criterion (AUC 0.63; p<0.001). Model calibration was robust, and the score retained full interpretability and clinical applicability. Conclusion An interpretable ECG-based machine-learning score accurately differentiates cardiac amyloidosis from hypertrophic cardiomyopathy in patients with ventricular hypertrophy, outperforming traditional voltage-based criteria. This pragmatic approach may facilitate earlier diagnosis and optimise referral for advanced imaging, particularly in resource-limited settings.