Phenotypic clustering of repaired Tetralogy of Fallot using unsupervised machine learning

Repaired Tetralogy of Fallot (rTOF), a complex congenital heart disease, exhibits substantial clinical heterogeneity. Accurate prediction of disease progression and tailored patient management remain elusive. We aimed to categorize rTOF patients into distinct phenotypes based on clinical variables and variables obtained from cardiac magnetic resonance (CMR) imaging.

A retrospective observational cohort study of rTOF patients with at least two CMR assessments between 2005 and 2022 was performed. From patient records, clinical variables, CMR measurements, and electrocardiogram data were collected and processed. Baseline and follow-up variables between subsequent CMR studies were used to assess both inter- and intrapatient disease heterogeneity. Subsequently, unsupervised machine learning was performed, involving dimensionality reduction using principal component analysis and K-means clustering to identify different phenotypic clusters.

In total, 155 patients (54.2 % male, median 14.9 years) were included and followed for a median duration of 9.9 years. A total of 459 CMR studies were included in analysis for the identification of phenotypic clusters. Following analysis, we identified four distinct rTOF phenotypes: (1) stable/slow deteriorating, (2) deteriorating, structural remodeling, (3) deteriorated indicated for pulmonary valve replacement, and lastly (4) younger patients with coexisting anomalies. These phenotypes exhibited differential clinical profiles (p < 0.01), cardiac remodeling patterns (p < 0.01), and intervention rates (p < 0.01).

Unsupervised machine learning analysis unveiled four discrete phenotypes within the rTOF population, elucidating the substantial disease heterogeneity on both a population- and patient-level. Our study underscores the potential of unsupervised machine learning as a valuable tool for characterizing complex congenital heart disease and potentially tailoring interventions.