JMIR Publications

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 (PRIME-MoE, Postoperative Risk Modeling Expert network) to provide dynamic, surgery-specific postoperative mortality risk prediction. The primary objective was to improve prediction accuracy, generalizability, and interpretability across heterogeneous surgical populations and institutions.

Methods:

We retrospectively analyzed 173,990 general anesthesia surgeries from two large-scale perioperative databases: INSPIRE (n = 99,735) and HUAXI (n = 74,255). The PRIME-MoE model integrates static preoperative features with full intraoperative multivariate time-series data through a sparse gating mechanism that activates surgery-specific expert networks. Model performance was compared with logistic regression, XGBoost, and LSTM baselines using AUROC, AUPRC, and subgroup analyses across surgical sites and ASA classifications. Integrated Gradients (IG) were used to interpret key predictors of postoperative mortality.

Results:

PRIME-MoE achieved AUROCs of 93.8 (95% CI: 92.3–96.0) on the INSPIRE database and 90.7 (87.1–94.8) on the HUAXI database, outperforming all baseline models. Cross-database validation demonstrated good generalization, with an AUROC of 90.8 (88.6–93.1) when trained on HUAXI and tested on INSPIRE, and 89.4 (84.5–93.0) in the reverse setting. Subgroup analyses showed robust and balanced performance across surgical sites and ASA physical status categories, with the greatest improvement observed in cardiovascular and vascular procedures. Integrated Gradients analysis identified heart rate, arterial pressure, oxygenation, and body temperature as key contributors to postoperative mortality risk, consistent with established clinical understanding.

Conclusions:

PRIME-MoE provides an interpretable, surgery-adaptive framework for postoperative mortality prediction that outperforms conventional scoring systems and deep learning baselines. By modeling heterogeneous intraoperative dynamics through specialized experts, it offers a scalable, generalizable, and clinically transparent decision-support tool for perioperative care.