JMIR Publications

ABSTRACT

Background:

Surgical site infections (SSI) remain the leading healthcare associated infections, making early detection critical for improving patient outcomes. While machine learning (ML) approaches have great potential for enhancing both diagnostic and prognostic predictive models for SSIs, the "black box" nature of complex models hinders their adoption in clinical decision-making. Explainable ML has emerged as a solution to provide necessary transparency, yet its application in SSI prediction remains limited.

Objective:

This review aims to provide an overview of the explainability operation of data-driven EML models for SSI diagnosis and prognosis from clinical care-oriented application perspectives, identifying key translational gaps between algorithmic development and real-world implementation.

Methods:

This scoping review was prospectively registered in PROSPERO (CRD420251124760) and adheres to the PRISMA-ScR guidelines. We searched PubMed, Web of Science, Scopus and Embase databases between January 2010 and August 2025. Inclusion criteria were peer-reviewed studies that developed data-driven explainable machine learning models for surgical site infection predictions. Data items charted from studies included bibliographic, model characteristics, predictor characteristics, explainer tools and corresponding results. Data were processed into categories that inductively emerged from the data and were synthesized using descriptive statistics.

Results:

Overall, 877 records were retrieved, and 38 articles that develop 48 machine learning models were included. Approximately 60% of these models were post-hoc explainable models. Regarding explanatory methods, 39 models provided ‘feature importance/contribution’ values with diverse calculation processes, while 4 models focused on the points that the model highly attended. The predictors used in all models were classified into 19 categories. Prognostic models tend to leverage preoperative and intraoperative information, specifically demographics and surgery parameters, with only 4 incorporating perioperative or early postoperative predictors. In contrast, most diagnostic models rely heavily on postoperative factors, where administrative codes often predominate. While these frequently utilized predictors, such as individual characteristics, health and lifestyle, comorbidities, and surgical details, often show higher contributions, some modifiable factors, such as laboratory results, admission management protocols, and specific preoperative preparations, were rarely identified as top-tier predictors.

Conclusions:

Compared with previous reviews that primarily focused on algorithm development or predictive performance, this review places greater emphasis on how the explainability is operationalized in explainable ML models for SSI. Overall, the depth and width of existing explanations are confined to the statistical significance and exhibit limited alignment with causal connections. The explanations mainly focus on post-hoc analysis on variable categories, neglecting the importance of timestamps, and the contextual grounding of explanations is often ambiguous. To generate clinically actionable and comprehensible explanations for medical practitioners, future work should develop explainer tools for causal channel or temporal attribution and formulate a detailed illustration framework of explanation contexts and contents. Clinical Trial: PROSPERO (CRD420251124760)