Development and external validation of a lightweight, explainable clinical decision support system for personalized perioperative risk stratification using electronic health records

Abstract

Abstract Late-preoperative risk stratification after final surgical scheduling may support perioperative risk communication, monitoring escalation, and resource coordination, yet many established calculators are difficult to automate within structured EHR workflows. We developed and externally validated LiteSurgFormer, a lightweight explainable MLP–attention risk-stratification model, in a retrospective multicenter cohort of 58,630 adult grade III–IV surgical patients treated at six Chinese sites from 2021 to 2024. Predictions were anchored after final operating-room schedule confirmation and primary surgical-team assignment but before incision, using 25 structured patient, disease, procedure, and surgical-team/scheduling predictors objectively available at that timestamp; intraoperative and postoperative variables were excluded. Zhejiang sites were used for model development and temporal internal validation, whereas a geographically external Xinjiang Alar affiliated-center cohort was isolated for final validation without refitting, recalibration, or domain adaptation. The primary endpoint was a 30-day composite of clinically significant postoperative adverse events. In external validation ( n  = 9772; 1308 events), LiteSurgFormer achieved an AUC of 0.912, Brier score of 0.070, and Hosmer–Lemeshow P  = 0.227, with higher decision-curve net benefit than machine-learning comparators and implementable clinical baselines across clinically relevant threshold probabilities. It outperformed ASA-only logistic regression (AUC 0.656) and a 21-predictor core-clinical logistic regression benchmark (AUC 0.895). In the ≥ 20% predicted-risk stratum, observed and predicted risks were closely aligned (48.5% vs. 48.0%; calibration slope 1.036; Hosmer–Lemeshow P  = 0.422). Discriminative performance was retained after removing surgical-team variables (AUC 0.896) or clinician-derived composite scores (AUC 0.886). These findings demonstrate retrospectively that LiteSurgFormer provides workflow-aligned risk stratification in comparable structured EHR settings, but portability to administratively unrelated health systems and clinical effectiveness after active deployment remain unproven. Clinical trial number Not applicable.