17 Artificial intelligence for the detection of surgical site infection on wound images. A systematic review and meta-analysis

Aim Surgical site infection (SSI) is one of the most common post-operative complications and is associated with significant morbidity and mortality. This study aims to provide an overview of whether artificial intelligence can accurately identify a surgical site infection on postoperative wound images. Methods The PubMed, Embase, CINAHL, Cochrane Library, Scopus, Web of Science, IEEE Explore, ArXiv and ACM Digital Library databases were searched up to November 12th 2024, and eligible studies were included. Risk of bias was assessed in the different studies using Prediction Model Risk of Bias Assessment Tool (PROBAST). A systematic review was conducted by combining data from the separate trials and analysing multiple area under the receiver operating curve (AUC) values. A meta-analysis to combine different AUC values was performed. Results We identified 12 eligible studies for inclusion. Out of 15 941 images from nine different surgical subspecialties, 1995 (12.5%) images were suspected for SSI. SSI labelling was either done by physical exam by a general practitioner or photographic data was judged by clinical experts. All included studies utilized both traditional machine learning or deep learning approaches. Overall risk of bias varied from low to high. Nine studies were found eligible for inclusion in the meta-analysis and analysis of pooled data from the nine included studies demonstrated a combined AUC of 0.92 (95% c.i. 0.68–0.99). Conclusion The use of AI models for SSI detection on images could prove beneficial in automating post-discharge surgical wound follow-up, possibly reducing the need for face to face evaluation and, in consequence, costs. However, current models do not yet meet the performance standard on sensitivity needed for them to be useful in postoperative wound monitoring. In this context, international data sharing and collaboration is key, both to increase generalizability and to improve model robustness with more images. Moreover, future studies should strive to incorporate external validation to evaluate model performance.