SP1.8 - Reducing the burden to deliver remote postoperative surveillance using machine learning for surgical-site infection (SSI)

Abstract Aims Remote postoperative wound (RPW) surveillance has been shown to improve the delivery of care and minimise the harm to patients from surgical-site infection (SSI). The burden for healthcare staff to deliver is a major barrier to implementation, therefore we aimed to develop novel methods for automated assessment of RPW surveillance to reduce this. Methods This was a secondary analysis of two interventional studies on RPW surveillance: “Tracking wound infection with smartphone technology” (TWIST) and “ImplementatioN of Remote Surgical wOund Assessment During the coviD-19 pandEmic” (INROADE). Adult general surgery patients could submit images of their surgical wound(s), and patient-reported outcome measures (PROMs) of SSI for 30-days postoperatively. A multi-input neural network (MNN) was developed to predict a clinical diagnosis of SSI within 48h based on symptoms and wound images. Performance was evaluated using area under the curve (AUC) and externally validated. Results There were 1540 submissions containing PROMs (48h SSI rate = 3.7%, n=57) and 2615 images (48h SSI rate = 3.1%, n=82). The MNN exceeded performance of the component models within the external validation subset (AUC: 0.94, 95% CI: 0.89-0.99) and remained equivalent to clinician triage (AUC: 0.92, 95% CI: 0.90-0.94). Usage to screen out “low-risk” responses prior to clinical triage was estimated to reduce the staff-time required to deliver (6.3h vs 25.5h). Conclusion Automated assessment can be successfully deployed within RPW surveillance pathways to reduce the burden on staff to deliver without compromising care, and allow resources to be appropriately directed to those at greatest risk of SSI.