Development and Validation of a Mixture-of-Experts–Based Model for Postoperative In-Hospital Mortality Risk Assessment: Multicenter Retrospective Study

Abstract Background Postoperative mortality remains a key indicator of surgical safety and quality. Traditional risk models based on static preoperative or intraoperative variables often fail to capture dynamic physiological changes during surgery and generalize poorly across surgical types. To address these limitations, we developed and validated a deep learning framework capable of learning heterogeneous temporal patterns from intraoperative signals while maintaining interpretability and adaptability. Objective This study aimed to develop a mixture-of-experts–based model (Postoperative Risk Modeling Expert network, PRIME-MoE) to provide dynamic, surgery-adaptive postoperative in-hospital mortality risk prediction. The primary objective was to improve prediction accuracy, cross-database robustness, and interpretability across heterogeneous surgical populations and institutions. Methods We retrospectively analyzed 173,990 surgical cases from the INSPIRE (Informative Surgical Patient Dataset for Innovative Research Environment) and HUAXI databases. PRIME-MoE used 4 one-layer unidirectional long short-term memory experts with dense softmax routing based on static surgical context. We evaluated discrimination, calibration, threshold-specific operating points, parameter count, graphics processing unit inference latency, routing patterns, and a sensitivity analysis that truncated the final 30 minutes of each operation without retraining. Results PRIME-MoE achieved internal-test area under the receiver operating characteristic curves (AUROCs) of 93.8 (95% CI 92.3‐96.0) in INSPIRE and 90.7 (95% CI 87.1‐94.8) in HUAXI. Reciprocal evaluation retained discrimination (AUROCs 90.8 and 87.8) but showed imperfect calibration (Brier scores 0.035‐0.049; calibration slopes 0.28‐0.49), supporting site-specific recalibration. Graphics processing unit inference required a median of 11.4 ms (9.1-18.0; 3.3-45.3) per case. Without retraining, scoring 30 minutes before the recorded end of surgery retained AUROCs of 93.1 in INSPIRE and 90.2 in HUAXI. Conclusions PRIME-MoE provides an interpretable framework for postoperative in-hospital mortality prediction and retained discrimination in reciprocal evaluation across 2 heterogeneous perioperative databases. The 30-minute analysis supports near-end-of-surgery scoring, but calibration was insufficient for direct deployment. Prospective validation, local recalibration, threshold selection, and workflow evaluation are required before clinical use.