Towards Digital Twins in Prostate Cancer: A Mixture-of-Experts Framework for Multitask Prognostics in Hospital Admissions

Abstract

Early risk prediction is essential for hospitalized prostate cancer (PCa) patients, who face acute events, such as mortality, ICU transfer, AKI (acute kidney injury), ED30 (unplanned 30-day Emergency Department revisit), and prolonged LOS (length of stay). We developed an MMoE (Multitask Mixture-of-Experts) model that jointly predicts these outcomes from the features of the multimodal EHR (Electronic Health Records) in MIMIC-IV (3956 admissions; 2497 patients). A configuration with six experts delivered consistent gains over strong single-task baselines. On the held-out test set, the MMoE improved rare-event detection (mortality AUPRC (Area Under the Precision-Recall Curve) of 0.163 vs. 0.091, +79%) and modestly boosted ED30 discrimination (AUROC (Area Under the Receiver Operating Characteristic Curve) ≈0.66 with leakage-safe ClinicalBERT fusion) while maintaining competitive ICU and AKI performance. Expert-routing diagnostics (top-1 shares, entropy, and task-dead counts) revealed clinically coherent specialization (e.g., renal signals for AKI), supporting interpretability. An efficiency log showed that the model is compact and deployable (∼85 k parameters, 0.34 MB; 0.027 s/sample); it replaced five single-task predictors with a single forward pass. Overall, the MMoE offered a practical balance of accuracy, calibrated probabilities, and readable routing for the prognostic layer of digital-twin pipelines in oncology.

Ähnliche Arbeiten