Machine Learning Reveals Novel Pediatric Heart Failure Phenotypes with Distinct Mortality and Hospitalization Outcomes

TL;DR

Machine learning identifies three distinct pediatric heart failure profiles with unique risks and treatment needs, suggesting better management strategies.

Contribution

Novel pediatric heart failure phenotypes are identified using machine learning, revealing distinct clinical outcomes and therapeutic implications.

Findings

Three distinct pediatric heart failure phenotypes were identified with unique clinical and outcome characteristics.

Cluster 1 showed frequent hospitalizations and beta-blocker use, while Cluster 2 had high mortality and prolonged stays.

Cluster 3 exhibited fulminant myocarditis with bimodal outcomes and high IVIG use.

Abstract

Background: Pediatric heart failure (PHF) is a heterogeneous syndrome with high morbidity, but existing classification systems inadequately capture its developmental and pathophysiological complexity due to reliance on adult-centric parameters. Using machine learning, we aimed to identify clinically distinct PHF phenotypes with unique outcomes and therapeutic implications. Methods: In this multicenter retrospective study, we analyzed 2903 consecutive PHF patients (≤18 years) from 30 Chinese tertiary centers from 20 provinces (2013–2022). Unsupervised machine learning (k-means clustering with PCA) evaluated 99 clinical, biomarker, and echocardiographic variables to derive phenotypes, which were compared for mortality, hospitalization, and treatment responses. Results: Three phenotypically distinct clusters emerged. Cluster 1 (Chronic Hypertensive and Cardiorenal Profile, 30.1%) predominantly affected older children (78%) with hypertension (54.4%), renal dysfunction (creatinine 45.8 μmol/L), and ventricular tachycardia (53.8%). This cluster showed the lowest in-hospital mortality (2.5%) but frequent 7–14 day hospitalizations (35.8%) and the highest beta-blocker use (54.5%). Cluster 2 (Preterm and CHD-Associated HF, 43.4%) comprised preterm infants (71.4%) with congenital heart disease (72.2%) and preserved LVEF (67%), demonstrating the highest mortality (5.1%) and prolonged stays (>30 days: 10.6%) with predominant diuretic (40.6%) and antibiotic use (54.3%). Cluster 3 (Fulminant Myocarditis Profile, 26.5%) exhibited cardiogenic shock with severely reduced LVEF (33%) and elevated BNP (3234 pg/mL), showing bimodal outcomes (4.8% LOS < 3 days vs. 32.2% LOS 15–30 days) and the highest IVIG utilization (46.5%) with intermediate mortality (3.8%). The majority of between-group differences were statistically significant (p < 0.001). Conclusions: Machine learning identified three PHF phenotypes with distinct in-hospital risk profiles and therapeutic implications, challenging current classification systems. These findings highlight the potential for phenotype-specific management strategies and provide a rationale for future research into arrhythmia prevention in hypertensive profiles and early immunomodulation in fulminant myocarditis, while highlighting the need for specialized care pathways for preterm/CHD patients. Prospective validation is warranted to translate this framework into clinical practice.

Figures