Advancing risk prediction in sickle cell disease through multi-task deep learning of mortality and intermediate risk factors

Abstract Sickle cell disease (SCD) displays a wide spectrum of clinical severity that becomes more apparent with age. Transplantation and genetic therapies are potentially curative, they carry high risks, underscoring the need for the ability to identify patients at the greatest risk of early mortality and significant morbidity. Survival analysis, however, is often complicated by censoring due to limited study duration, loss to follow-up, competing risks and does not capture multi-dimensional interactions. Longitudinal progression of risk factor, such as lab values or echocardiographic (Echo) parameters, offers prognostic signals but is seldom integrated with survival models, leading to inconsistencies. Here, we developed a deep learning multi-task model (Multi-Task DeepHit) that uses baseline covariates from a well-characterized SCD patient cohort to predict mortality and risk factor trajectories in adults with SCD. The model employs masked attention and hard parameter sharing to improve robustness and capture shared information. The study involved 598 SCD patients (median age 33.5 [IQR 25–45], 304 (51%) female) who underwent baseline Echo and lab tests in prospective cohort studies (NCT00011648 and NCT00081523) at the NIH Clinical Center from 2006 to 2025. Over a median follow-up of 7.1 years, 218 deaths (36.5%) were observed. We analyzed 68 baseline covariates (demographics, vitals, genotype, labs, Echo) and 12 longitudinal functional, hepatic, renal, and cardiac parameters within 3 years after baseline to predict 15-yr all-cause mortality. Covariates with <30% missing data were imputed by random survival forest (RSF). We evaluated our Multi-Task DeepHit model against two other methods - two-step statistical machine learning (TSML) and DeepHit - using bootstrap-corrected C-statistics and integrated Brier scores (IBS). Patients were stratified into high (≥10%, N=100) and low-risk (<10%, N=498) groups for 5-year mortality prediction. SHapley Additive exPlanations (SHAP) analysis identified key predictors and Wilcoxon rank-sum tests compared key predictor distributions. Sensitivity analyses evaluated scenarios with incomplete baseline data. Conformal prediction assessed individual uncertainty in yearly mortality risk. The Multi-Task DeepHit significantly outperformed both the TSML and DeepHit models, with the highest bootstrap-corrected C-statistic of 0.8617 and the lowest 5-yr IBS of 0.0300, compared with C-statistics/IBS of 0.7399/0.0582 for TSML and 0.8383/0.0503 for DeepHit. When simulating 10%, 20%, and 30% missing data, our model still maintained superior C-statistics compared to TSML (0.8284, 0.7883 and 0.7574 vs. 0.7399) and IBS (0.0401, 0.0492 and 0.0555 vs. 0.0582). The 95% conformal prediction intervals were well-calibrated over years 1–10, achieving an average empirical coverage of 95.6% (SD 3.2%), closely aligning with the nominal level. SHAP analysis identified 5 key predictors for 5-yr mortality: reticulocyte count, alkaline phosphatase (ALP), right atrial (RA) pressure, tricuspid regurgitation peak velocity (TRV), and RA area. RA pressure was the top predictor in the high-risk group while reticulocyte count ranked first in the low-risk group. Wilcoxon rank-sum tests revealed that RA pressure, TRV, RA area and ALP levels were significantly higher in high-risk patients compared to low-risk patients (all p < 0.001). Reticulocyte count showed no significant difference (p = 0.34), suggesting its importance as a predictor may be due to nonlinear or interaction effects captured by the model rather than direct differences in distribution. We evaluated the utility of our model in predicting 5-yr mortality for individual patients. For example, Patient #1 had a predicted 5-yr mortality of 97.91%, and death occurred at 4.1 years. In contrast, Patient #2 had a predicted 5-yr mortality of 0.57%, and death occurred after 11 years. The Multi-Task DeepHit accurately predicts mortality risk in patients with SCD, demonstrating strong performance even with missing baseline data. It provides reliable uncertainty estimates and effectively models both mortality and changes in risk factors over time. By using SHAP values for individual predictions, the model identifies key predictors of mortality, highlighting the potential of multi-task deep learning to enhance risk stratification and personalized care in SCD. Future studies should focus on external validation and assessing its usefulness as a decision tool in the clinical setting.