Perioperative Data Improve Postoperative Complication Prediction
Key Takeaways
- Among adults undergoing nonemergency noncardiac surgery under general anesthesia at three affiliated Japanese hospitals, the defined 30-day complication composite occurred in 10.9%.
- Operative and early postoperative inputs improved hospital-held-out discrimination over preoperative-only inputs across three algorithms.
- Expanded predictions tended to overestimate event risk, and calibration varied by hospital.
- Exploratory decision curves generally showed greater net benefit for expanded models, although the curves overlapped substantially.
In a JMIR Perioperative Medicine study, researchers retrospectively analyzed 35,702 adult inpatients undergoing nonemergency noncardiac surgery under general anesthesia at three affiliated Nippon Medical School hospitals in Japan from January 2021 through March 2025; emergency and cardiovascular operations were excluded. The 30-day composite included in-hospital death, unplanned reoperation, invasive ventilation for at least 48 hours or reintubation within 48 hours, newly initiated renal replacement therapy, and infection defined by at least 72 hours of nonprophylactic systemic antibiotics. Logistic regression, extreme gradient boosting, and light gradient boosting machine (LightGBM) each compared 36 preoperative predictors with those predictors plus seven additional measures: operative duration, red-cell, fresh frozen plasma, and platelet transfusion volumes, fluid balance, and heart rate and mean arterial pressure six hours after surgery. The expanded prediction was therefore unavailable before surgery. For primary leave-one-facility-out validation, investigators developed, imputed, tuned, and isotonic-calibrated models using only the other two hospitals before testing each held-out hospital and pooling predictions.
The composite occurred in 3,884 patients, with antibiotic-defined infection accounting for most events. Across the three algorithms, pooled held-out area under the receiver operating characteristic curve (AUROC) was 0.784–0.791 with preoperative inputs versus 0.810–0.817 with expanded inputs. Area under the precision–recall curve (PR-AUC) was 0.356–0.361 versus 0.405–0.413, respectively. Secondary random hold-out and later temporal validation showed directionally similar discrimination gains.
Overall Brier performance also improved with expanded inputs, but pooled perioperative calibration slopes were 0.875–0.902 and intercepts were −0.367 to −0.222. Predictions were overly extreme, generally overestimated event risk, and differed in calibration among individual hospitals. Exploratory decision curves generally favored expanded models despite substantial overlap; a positive prediction represented prioritized postoperative reassessment, not a prespecified intervention. An analysis excluding infection retained better LightGBM discrimination but poor calibration.
All facilities belonged to one Japanese academic network, so holding out a hospital did not establish transferability to unrelated health systems or North American settings. Retrospective model evaluation did not test whether using predictions changes patient outcomes. The heterogeneous outcome was heavily influenced by infection identified through antibiotic administration rather than independently confirmed diagnoses. According to the investigators, local recalibration and prospective workflow assessment remain unresolved implementation needs; neither decision curves nor predictor contributions establish a treatment, deployment threshold, or causal relationship.
The researchers concluded that routinely recorded operative and early recovery inputs produced modest, consistent predictive gains over preoperative information. They noted that between-hospital calibration differences and the absence of prospective clinical-impact testing leave patient benefit undemonstrated.
Clinician Questions
Did the noncardiac surgery complication models outperform ASA-PS or the Revised Cardiac Risk Index?
Investigators compared perioperative and preoperative predictor sets within their own algorithms for adults undergoing nonemergency noncardiac surgery under general anesthesia. They did not directly compare performance against American Society of Anesthesiologists Physical Status (ASA-PS), the Revised Cardiac Risk Index, or a surgical risk calculator, so relative superiority over those tools was not established.
Were postoperative myocardial injury or other nonfatal cardiac events counted in the noncardiac surgery complication outcome?
The 30-day composite for adults undergoing nonemergency noncardiac surgery under general anesthesia excluded nonfatal cardiac complications. Postoperative myocardial injury could not be consistently identified because serial postoperative cardiac troponin testing was not routine across the three affiliated Japanese hospitals.
How well does the Japanese postoperative risk model apply to higher-risk surgical patients?
Most patients at the affiliated Japanese hospitals were classified as ASA-PS I or II. The authors identified applicability to higher-risk surgical populations as unresolved.
This content is intended for healthcare professionals for educational purposes. The article and its image were developed with AI assistance and reviewed by members of the ReachMD team prior to publication. Please refer to the original source linked for the complete and most up-to-date information.