Advancing mortality risk prediction in sickle cell disease through multi-task deep learning models

Patients with sickle cell disease (SCD) have highly variable clinical courses that become more heterogeneous over time. As curative therapies such as hematopoietic cell transplantation and gene therapy become more available but remain risky, clinicians need accurate, interpretable tools to identify patients at increased mortality risk and the organ-system complications driving it. We developed Multi-Task DeepHit, a neural-network survival model that uses baseline clinical data and learns from 3-year biomarker trajectories during model development. Once trained, it required only baseline data for mortality prediction. We studied 598 adults with SCD (median age 34 years; 51% female, 2006 - 2025) in one NHLBI cohort and externally validated the model in a separate, non-overlapping NHLBI cohort of 383 adults. Baseline inputs included 68 demographic, laboratory, vital-sign, and echocardiographic covariates. Twelve longitudinal biomarkers captured renal, hepatic, cardiopulmonary, and systemic function. Multi-Task DeepHit outperformed both a two-step statistical/machine-learning model and a survival-only DeepHit model, with better discrimination and lower 5-year prediction error in both cohorts. This improvement reflected true patient-specific trajectories, not added model complexity, as performance fell when trajectories were shuffled. The strongest contributors to predicted mortality were reticulocyte percentage, alkaline phosphatase, right atrial pressure, tricuspid regurgitation velocity, and right atrial area. Lower reticulocyte percentage contributed to higher predicted risk, although this pattern depended on other biomarkers. By leveraging longitudinal biomarker trajectories, Multi-Task DeepHit improves mortality prediction in adults with SCD, supports individualized risk assessment, and identifies patients who may benefit from closer surveillance or more intensive disease-modifying or curative therapies.