Explainable machine learning predictive models for surgical site infections : scoping review

Abstract

Background: Surgical site infections (SSIs) remain a major cause of health care–associated infections, and early prediction is essential for improving patient outcomes. Machine learning (ML) has shown potential for SSI prediction; however, clinical implementation requires models that are both accurate and explainable. Despite recent progress in explainable ML, its clinical application to SSI prediction remains limited.

Objective: This study aimed to map explainable ML models for SSI prediction from a clinical perspective and examine their use of structured and unstructured data across the dimensions of data, methodology, and explanation output.

Methods: We conducted a scoping review following PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) and Joanna Briggs Institute (JBI) guidance, and registered the protocol in PROSPERO. Six databases were searched for eligible studies published from January 2010 onward, without language restrictions. The search was conducted on August 9, 2025, and updated on July 14, 2026. We included studies that developed or validated an explainable ML model for predicting SSI in adults. Two reviewers (RS and YL) independently screened studies and extracted data. Findings were narratively synthesized and presented in evidence maps. Methodological quality was assessed using the PROBAST+AI (Prediction model Risk Of Bias Assessment Tool for prediction models using regression or artificial intelligence) methods.

Results: Overall, 77 studies reporting 98 ML models were included. Most models were prognostic (72/98, 73.5%), whereas 26 focused on postoperative SSI diagnosis. A total of 81.8% (63/77) of studies addressed a single surgical specialty, most commonly gastrointestinal surgery (27/63, 42.9%). Overall, 51.9% (40/77) of studies addressed composite SSI predictions. Among all models, % (40/98) were black-box models explained by post hoc methods; SHAP combined with ensemble learning was the leading approach (18/40, 45%). Regression models accounted for half of the inherently interpretable models, interpreted using coefficients. Prognostic models commonly included health and lifestyle (60/72, 83.3%), individual characteristics, and surgical process details (both 57/72, 79.2%); health and lifestyle factors were most frequently important across SSI types.Diagnostic models commonly included surgical process details (13/26, 50%), administrative codes, and individual characteristics (both 10/26, 38.5%). Key diagnostic predictors varied by SSI types: postoperative clinical interventions predominated for composite SSI; vital signs, postoperative interventions, and administrative codes for superficial SSI; postoperative recovery status for deep SSI; and vital signs for organ-space SSI.

Conclusions: Extending previous reviews focusing on model performance, this review mapped explainability methods and important features in SSI prediction, identifying recurring predictor patterns and substantial methodological heterogeneity across prognostic and diagnostic settings. Incomplete reporting of feature definitions and explanatory rationale, together with limited clinical relevance, constrained clinical interpretation and actionability. Clinician-informed reporting frameworks and validation of explanation fidelity and clinical relevance are needed to improve the trustworthiness and utility of SSI prediction models.