Echocardiographic Clustering by Machine Learning in Children with Early Surgically Corrected Congenital Heart Disease

Abstract

The research investigates the time-series clustering from echocardiographic data in children with surgically corrected congenital heart disease (CHD). In recent years, machine learning has been demonstrated to discover sophisticated latent patterns in medical data, yet relevant explainable applications in pediatric cardiology remain lacking. To address this issue, we propose an autoencoder-based architecture to model time-series data with interpretable results effectively. The proposed method outperforms the baseline models in terms of internal clustering metrics. The three clusters also show distinguished differences in patients' outcomes. Patients in Cluster 0 exhibit the poorest prognosis, with an approximate reoperation rate of 40\% within the initial six months following the index surgery. The data mining result can potentially facilitate clinicians to stratify patients' prognoses based on echocardiographic and clinical observations in the future.