An Attention-based Deep Learning Approach for Lifespan Assessment of Heart Failure Risk Among Patients with Congenital Heart Disease

Objective—Congenital heart disease (CHD) presents persistent challenges and risks, including long-term comorbidities such as heart failure (HF), necessitating precise delivery of care. This study aims to develop a machine learning method for assessing the lifespan HF risk trajectories by incorporating the comprehensive medical histories of patients with CHD. Methods—We developed hART (heart failure Attentive Risk Trajectory), a deep-learning model to predict HF trajectories in CHD patients. hART is designed to capture the contextual relationships between medical events within a patient’s history. Specifically, it uses masked self-attention mechanisms to focus on the most relevant segments of the past medical events while not peaking ahead of the future events. To demonstrate the utility of hART, we used a retrospective cohort containing healthcare administrative data from the Quebec CHD database (137,493 patients, 35-year follow-up). We evaluated hART’s performance by area under the receiver operating characteristic (AUROC) curve and area under the precision-recall curve (AUPRC) in predicting future HF compared to the state-of-the-art methods. We further evaluated the effectiveness of hART by examining the differences in HF risk trajectory for patient subgroups, including those with genetic syndrome and severe CHD lesions, as well as patients who died at different ages. Additionally, we computed individualized trajectories and extracted attention weights to assess how specific medical events contribute to rising predicted HF risk. Finally, we extended hART by developing hART-Generative Pre-trained Transformer (GPT), which is pre-trained to learn the clinical language of the diagnoses and comorbidities conditions across all patients and then fine-tuned to more accurately predict HF compared to the baseline hART that was trained to predict HF from scratch.Results-hART outperformed existing methods, achieving an AUROC of 0.967 and an AUPRC of 0.282 for HF risk prediction. The analysis of computed HF trajectories across different populations revealed that patients with severe CHD lesion consistently exhibited elevated HF risks throughout their lifespan. This indicates the potential for the use of hART for effective risk stratification. Patients with the genetic syndrome of CHD exhibited elevated HF risks until the age of 50. Notably, we found a decrease in the impact of the birth condition on long-term risk, emphasizes how hART recognizes the significance of birth conditions and their varying impact on HF risk over different lifespans. Moreover, our study showcased how hART captured the importance of the timing of medical events, such as surgery. By analyzing the HF trajectory of individual patients, hART demonstrated that the timing of arrhythmic surgery had varying impacts on lifespan HF risk, as we demonstrated that arrhythmic surgery performed at a younger age had minimal long-term effects on HF risk, while surgeries during adulthood had a significant lasting impact. This underscores the model’s ability to consider the context and timing of medical events. The hART-GPT model demonstrated superior accuracy in predicting comorbidities associated with HF, such as stroke, infective endocarditis, sepsis, MI, and acute kidney disease. Furthermore, after fine-tuning, it demonstrated slightly improved HF prediction over hART. Conclusions-This study demonstrated that attention-based deep learning models can accurately assess lifelong HF risk in patients with CHD. This study developed a model that accurately captures both long—and short-range dependencies in patient histories while offering enhanced interpretability for clinicians through the inclusion of HF trajectories and attention matrices. The interpretable disease trajectories provided by our model have the potential to enable clinicians to identify high-risk individuals, optimize intervention timing, and assess the long-term impact of comorbidities in CHD patients

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.