Identifying patient-reported outcome patterns in different congenital heart defects using unsupervised machine learning

Abstract Background Adults with congenital heart disease (CHD) represent a growing and clinically heterogeneous population with diverse symptom burden, functional status, and quality of life. Patient-reported outcomes (PROs) capture this variability, yet average scores obscure meaningful subgroups and hinder identification of "vulnerable" patient subgroups. Distinct PRO profiles could inform patients stratification and clinical care. Purpose (i) To identify distinct PRO profiles of adults with CHD, and (ii) to determine their distribution across different heart defects (lesion types). Methods The ‘Assessment of Patterns of Patient-Reported Outcomes in Adults with Congenital Heart disease – International Study II’ (APPROACH-IS II) enrolled 8,415 adults with CHD across 52 centers in 32 countries. PROs included perceived health status (RAND-12 Health Survey; Linear Analog Scale Health Status); depression (Patient Health Questionnaire-8); anxiety (General Anxiety Disorder-7); and quality of life (Linear Analog Scale Quality of Life). Unsupervised machine learning methods, including Uniform Manifold Approximation and Projection (UMAP) and K-means clustering, were applied to identify and visualise PRO-based patterns. χ² test compared PRO profiles across different heart defects. Results Four clusters were identified. Cluster 1 included patients with a high overall burden and globally poor PROs (12.95%). Cluster 2 represented globally good PROs (46.81%). Cluster 3 included patients with relatively good psychological health but impaired physical functioning and reduced quality of life (21.27%). Cluster 4 included patients with relatively good physical functioning and quality of life but poor psychological health (18.97%) (Fig 1). Heart defects differed significantly across PRO clusters (χ²= 253.69; df = 69; p 0.001) (Fig 2). Cyanotic defects; moderate and large unrepaired secundum ASD; and Ebstein anomaly was overrepresented in Cluster 1 and 3, characterizing overall poor PRO or limitations in physical functioning. Still, one third of these patients were in Cluster 2. Patients with aortic valve disease, coarctation of the aorta, and isolated small VSD were overrepresented in Cluster 2 and 4, showing that these patients were doing well overall or in terms of physical functioning (Fig 2). Some complex defects were more common in cluster 1, while some simple/moderate defects were more frequently observed in clusters 2 and 4. Conclusions This study uncover four distinct PRO profiles in adults with CHD. Certain subtypes of CHD were associated with a worse global profile (Cluster 1), positive overall profile (Cluster 2), positive psychological well-being but worse physical functioning and quality of life (Cluster 3), and poor psychological well-being but better physical functioning and quality of life (Cluster 4). Heart defect complexity also varied across PRO clusters, with simple/moderate defects tending to be more likely to belong to Cluster 2 and 4.