Image-Based Algorithms for Remote Surgical Site Infection Diagnosis in Rural Rwanda

Abstract

A key challenge in global health is the delivery of high-quality health care in low-resource settings. In sub-Saharan Africa, wound infections are physically and financially catastrophic for patients due to limited access to medical facilities, health professionals, and adequate follow-up care. Surgical Site Infections (SSIs) in rural Rwanda affect an average of 12.4% of women undergoing cesarean deliveries and significantly contribute to maternal morbidity and mortality. A key challenge to SSI monitoring is that cesarean-associated SSIs often develop after hospital discharge and traveling back to a facility for follow-up care is burdensome for women. However, the widespread availability of mobile health technology among community health workers (CHWs) in rural regions presents a new opportunity for SSI care. In this thesis, we propose a CHW-led mobile health system for automated, home-based diagnosis of surgical site infections for mothers after C-section. The mobile health system takes a smartphone image of the wound site, uses a real-time computer vision algorithm for image preprocessing, and deploys an AI-based predictive model for image classification. The methodology eliminates the need for tedious manual preprocessing and the variability in the collected image data. The utility of our method is demonstrated by testing on two image datasets -- visible and thermal -- of 621 wound photos collected from women at Kirehe District Hospital in Rwanda. The final model trained on visible images had an AUC=0.86 (sensitivity=0.83, specificity=0.75). The final model trained on the thermal images had an AUC=0.89 (sensitivity=0.94, specificity=0.81). Overall, we demonstrate the first smartphone-based methodology for surgical site infection diagnosis. Further, the thermal model provides a new opportunity to help identify wound infections without dependence on skin color. We believe this new diagnostic tool will reduce barriers to accessing post-partum follow-up care, and takes a promising first step toward improving maternal health in low-resource settings. Further, we hope this advancement paves the way for the practical implementation of image-based machine learning models and represents an important milestone for the adoption of artificial intelligence for global health.