Abstract Allogeneic hematopoietic stem cell transplantation remains critical for treating high-risk hematological malignancies like leukemia. Despite advances in treatment, early mortality remains clinically significant, with approximately 6–11% of patients dying within the first 100 days after transplantation. This highlights the need for dynamic monitoring strategies beyond static pre-transplant risk assessments. This study introduces a novel, interpretable, real-time proof-of-concept monitoring framework that continuously assesses individualized mortality risk during treatment using routinely collected clinical data. The framework employs deep learning models to predict seven-day mortality risk based on 22 laboratory parameters from the previous 14 days and five demographic features. The framework incorporates an explainability method that provides time-resolved insights into predictions, which can be aggregated across time intervals or patient groups for broader interpretation. We evaluated the approach on data from 891 patients treated at the University Hospital of Düsseldorf (UKD; 2004–2019), as well as on an independent cohort derived from the MIMIC-IV database. Our experiments demonstrate that the predicted mortality risk aligns with observed outcomes, achieving a patient-level AUROC of 0.95 in the primary (UKD) cohort. Preliminary expert evaluation suggests that the predictions and explanations are intuitive and clinically relevant, supporting awareness of complications and highlighting potential for timely intervention.
Rucks et al. (Thu,) studied this question.