A Super Learner AI model integrating electrocardiogram and echocardiography features accurately distinguished cardiac amyloidosis from other causes of left-ventricular hypertrophy with an AUC of 0.97.
Cross-Sectional (n=1,221)
Yes
Does an AI model integrating ECG and echocardiography accurately diagnose cardiac amyloidosis among patients with left-ventricular hypertrophy?
An AI model and simplified scoring system using routine ECG and echocardiography parameters can accurately distinguish cardiac amyloidosis from other causes of left ventricular hypertrophy.
Effect estimate: AUC 0.97 (95% CI 0.95-0.98)
Cardiac amyloidosis (CA) is an under-recognized cause of left-ventricular hypertrophy (LVH) that is often misclassified as hypertrophic cardiomyopathy (HCM) or hypertensive heart disease (HHD). We aimed to develop and externally validate an AI model using electrocardiograms (ECG) and echocardiography to distinguish CA from other LVH aetiologies. A retrospective, multicenter study was conducted, with a derivation cohort collected from PUMCH (290 CA patients, 215 HCM patients, and 160 HHD patients) and an external validation cohort recruited from 10 other hospitals across China (126 CA patients, 240 HCM patients, and 190 HHD patients). Twenty-eight clinical, ECG, and echocardiographic predictors were included, and we selected the top 7 important features by recursive feature elimination strategy. The Super Learner model combines predictions from 11 machine learning models, and model performance was evaluated via macro-AUC, accuracy, precision, F1 score, and etc. We developed a simplified scoring system to diagnose CA, and classifying patients into three groups based on CA probability to minimize misdiagnosis risk. Ranking by feature importance, the top seven features were included and used for model construction, including Sokolow–Lyon index, interventricular septal thickness, systolic blood pressure, left-ventricular posterior wall thickness, tricuspid annular plane systolic excursion, average E/e′, and left-ventricular ejection fraction. The Super Learner model, combining Extra Trees, Histogram-based Gradient Boosting, LightGBM, and Multi-Layer Perceptron, achieved the highest AUC of 0.97 (95% confidence interval CI: 0.95–0.98). In external validation, the super learner model achieved AUCs of 0.96 (95% CI: 0.95–0.98) for CA, 0.93 (0.91–0.95) for HCM, and 0.91 (0.89–0.94) for HHD. And the simplified scoring system also showed robust diagnostic performance (AUC 0.90, 95% CI 0.86–0.93). We developed an AI model integrating ECG and echocardiography that provides a clinically applicable and noninvasive framework for CA screening and diagnosis, and implementation through a WeChat-based screening program. However, its performance and generalizability should be further validated in larger prospective multicenter studies.
Zhang et al. (Thu,) conducted a cross-sectional in Cardiac amyloidosis (n=1,221). Super Learner AI model was evaluated on Diagnostic accuracy (AUC) for cardiac amyloidosis (AUC 0.97, 95% CI 0.95-0.98). A Super Learner AI model integrating electrocardiogram and echocardiography features accurately distinguished cardiac amyloidosis from other causes of left-ventricular hypertrophy with an AUC of 0.97.