Artificial intelligence (AI)-enhanced electrocardiography: a machine-learning model for differential diagnosis between hypertrophic cardiomyopathy, cardiac amyloidosis and Anderson-Fabry disease

Abstract Background The differential diagnosis between Hypertrophic Cardiomyopathy (HCM) and phenocopies is challenging and relies on a comprehensive approach, which includes clinical and instrumental data, as well as laboratory exams and genetic testing. In this setting, 12-lead electrocardiogram (ECG) represents a valuable tool, but although typical patterns have been described, their diagnostic ability to discriminate between different etiologies is limited and requires considerable clinical experience. However, a correct diagnosis is of pivotal importance, allowing the prompt initiation of phenotype-specific therapies and consequently improving patient survival. Purpose Aim of this study was to develop a machine-learning model based on human-based ECG features to support the differential diagnosis between HCM, Cardiac Amyloidosis (CA) and Anderson-Fabry Disease (AFD). Methods We retrospectively examined ECGs of 200 patients, 50 affected by HCM, 100 affected by CA (50 AL CA and 50 transthyretin CA) and 50 with a diagnosis of AFD cardiomyopathy. The patients were then divided in sinus rhythm (SR) subgroup (158 patients) and atrial fibrillation (AF) subgroup (42 patients). For each subgroup, patients were randomly split into a training (80%) and a testing cohort (20%), and a random forest (RF) model was trained on the training cohort, with hyperparameter tuning via 10-fold cross validation. The resulting model was then tested on the testing cohort to assess its predictive accuracy. Results Both RF models (SR and AF) showed good discriminative performance in this multiclass classification task on the testing cohort [SR: accuracy 67%; multiclass area under the curve (AUC) 0.797, 95% confidence interval (CI) 0.472 -0.827; AF: accuracy 75%; multiclass AUC 0.792, 95% CI 0.349 – 0.968]. Figure 1 and figure 2 report for each model the rank of the most important variables for class prediction. Conclusions An AI-based ECG interpretation based on human-derived ECG features can predict the etiology underlying an hypertrophic phenotype, helping to perform a correct diagnosis and consequently to set a specific therapeutic process.