hART: Deep learning-informed lifespan heart failure risk trajectories

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Heart failure (HF) presents persistent risks and long-term comorbidities, especially for patients with congenital heart disease (CHD) and its lifelong sequelae.
• Congenital heart disease (CHD) patients face unique challenges in managing heart failure (HF) risks throughout their lives.

Purpose of the Study:

• To develop hART (heart failure Attentive Risk Trajectory), a deep-learning model for predicting HF trajectories in CHD patients.
• To enhance the prediction of future heart failure risk by analyzing complex medical histories.

Main Methods:

• hART utilizes a masked self-attention mechanism to identify and prioritize relevant past medical events for HF risk prediction.
• The model captures contextual relationships between medical events to understand individual patient trajectories.
• A large cohort from the Quebec CHD database (137,493 patients, 35-year follow-up) was used to train and validate hART.

Main Results:

• hART achieved a 33% improvement in HF risk prediction, with an area under the precision-recall curve of 28%.
• Severe CHD lesions correlated with consistently elevated HF risks across the lifespan.
• Genetic syndromes were associated with elevated HF risks until age 50, with decreasing impact of birth condition over time.
• The timing of interventions like arrhythmia surgery significantly influenced long-term HF risk, with earlier surgeries having less impact.

Conclusions:

• hART effectively detects significant lifelong HF risks in CHD patients by analyzing both short- and long-range dependencies in medical histories.
• The model provides a novel approach to understanding and predicting the lifelong HF risk associated with congenital heart disease.