An Open-Source Analysis of Cardiomyopathy Using Machine Learning and Electrocardiograms

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Dilated cardiomyopathy (DCM) and hypertrophic cardiomyopathy (HCM) are leading causes of heart failure.
• Current diagnostic methods may not be readily accessible in all areas, highlighting the need for streamlined screening.
• Existing machine learning studies on cardiomyopathy often lack open-source data and direct comparisons between HCM and DCM subtypes.

Purpose of the Study:

• To develop and validate an open-source, ECG-based machine learning pipeline for differentiating between DCM and HCM.
• To identify distinct ECG features that characterize HCM, DCM, and subtypes like obstructive HCM (HOCM) and non-obstructive HCM (HNCM).
• To assess the feasibility of using ECG screening for early cardiomyopathy diagnosis.

Main Methods:

• Extracted standard and vectorcardiogram-derived (VCG) ECG features from the MIMIC-IV-ECG database.
• Utilized logistic regression (LR) and extreme gradient boosting (XGBoost) models with cross-validation on a cohort of 599 patients.
• Compared the models' performance in distinguishing HCM from DCM subtypes (DCM-I, DCM-NI) and HOCM from HNCM using AUC-ROC.

Main Results:

• LR models achieved high discrimination between HCM and both DCM subtypes (AUC-ROC 0.90-0.92).
• Differentiating HOCM from HNCM was more challenging (XGBoost AUC-ROC 0.81).
• Distinct ECG patterns were observed: DCM showed lower QRS amplitudes and right-posterior gradients, while HCM exhibited higher amplitudes and complex T-loops. HOCM displayed stronger leftward electrical activity.

Conclusions:

• An interpretable, open-access ECG analysis pipeline can effectively discriminate between cardiomyopathy types.
• Specific ECG features offer measurable separation between HCM and DCM, supporting their use in diagnostic screening.
• While differentiating obstructive from non-obstructive HCM via ECG remains challenging, the study provides a reproducible framework for future research.