AI-Integration of Multi-Day Images for Detecting Surgical Site Infections in Rural Rwanda

Purpose Surgical site infections (SSIs) are a major postoperative complication, particularly in resource-limited settings where routine follow-up is challenging. Our study explores the integration of multi-day imaging in deep learning models to improve the detection accuracy for SSIs. Materials and Methods We conducted a retrospective analysis of postoperative wound images from 733 women who underwent cesarean delivery at Kirehe District Hospital, Rwanda (Sept 2019–Mar 2020). Images were captured on postoperative day 3 (POD3) and day 11 (POD11). SSI diagnoses were made by trained general practitioners using CDC criteria. We compared five CNN-based models: (1) POD3 alone, (2) POD11 alone, (3) feature summation across days, (4) feature delta between days, and (5) direct integration via a Siamese network. Models were trained for 50 epochs with SMOTE, class weighting, and 5-fold cross-validation. Performance was evaluated by AUC, sensitivity, specificity, and bootstrapped 95% CIs. Results Our POD3-alone model had the lowest mean AUC (0.64), while the POD11-alone model achieved better performance (AUC: 0.85). Among the multi-day integration methods, summation of features (AUC: 0.76) and delta calculation (AUC: 0.67) underperformed compared to POD11-alone. Direct integration using a Siamese network achieved the highest AUC (0.89) and specificity (0.82). Conclusion Multi-day imaging integration enhances detection of SSIs by capturing progressive wound changes, offering advantages over single-day models. While direct integration provides the most accurate results, its computational intensity poses challenges for real-world deployment. Future work should focus on external validation, integration with thermal image data, ensemble strategies, and computational optimizations for deployment in low-resource settings.