ECG-based machine-learning approach to differentiate cardiac sarcoidosis from arrhythmogenic right ventricular cardiomyopathy

Cardiac sarcoidosis (CS) represents a rare and challenging condition with poor prognosis and complex diagnosis. This inflammatory disease is characterized by the formation of noncaseating granulomas in the myocardium and its clinical presentation is highly variable, mimicking other cardiac conditions as arrhythmogenic right ventricular cardiomyopathy (ARVC). ARVC is a hereditary disease characterized by fibrofatty infiltration that results in both left and right ventricular dysfunction which can be observed also in CS patients. Despite the critical need for accurate differentiation to ensure optimal patient management, current literature lacks comprehensive exploration of ECG-based features to characterize the two populations.

This study aims at developing a machine learning classifier based on ECG-based parameters, to distinguish between CS and ARVC patients.

The study analyzed 12-lead ECG recordings from 100 patients (50 with ARVC and 50 with CS), collecting 630 standard ECG measurements. These included amplitude, duration, area under the wave, and peak time for each wave component (P, Q, R, S, T), along with duration of the ECG complex components. The dataset was divided into training (70%) and test set (30%), using a patient stratified split. A two-step feature selection was perfomed, to assess features redundancy and relevance using Spearman coefficient with a correlation threshold equal to 0.90 and Mann-Whitney test (p<0.05), respectively. Three machine learning models were tested (decision tree, logistic regression and gradient boosting) on the selected features and their performances were evaluated in terms of accuracy, sensitivity and specificity. The Gini index was finally computed to determine features’ importance in patient classification. The five most relevant ECG-based features were used to develop five models.

Sixteen features were selected as non-redundant and relevant and used to train the three tested models. Gradient boosting reached the best performances with an accuracy, sensitivity and specificity of 0.867 on the test set, correctly classifying 13/15 patients of both ARVC and CS group. The Gini importance index identified following top features as relevant for the final group allocation including QRS duration, QTcFridericia, T-wave area (leads V1, V2, V3), R-wave amplitude (leads III and aVF), S-wave amplitude (lead III), T peak time (lead V1), and QRS area (lead aVL). The model based on QRS duration only reached an accuracy of 0.63, while adding one feature at time determined an increase in model accuracy reaching 0.87 with only five features.

This study demonstrates the potential of ECG-based machine learning models in distinguishing between ARVC and CS. This study demonstrates the potential of ECG-based features in distinguishing between ARVC and CS, opening new perspectives for accurate differential diagnosis in clinical settings.