Prediction of Diabetic Foot Ulcer Healing Potential Using Machine Learning

Abstract Body: Over 6 million Americans have diabetic foot ulcers (DFUs)1 which greatly increase the risk of morbidity and mortality2,3 as well as healthcare costs.4 Standard clinical practice evaluates wound healing in a temporal manner which can delay initiation of the most effective treatment. Prediction models for use in chronic wound treatment have been explored,5 but remain very limited as more work is needed to delineate the features most important in building algorithms. Using data from a clinical study, a multi-spectral imaging (MSI) device and machine learning were used to generate a model to predict the probability of DFU healing within 30 days of standard wound care therapy. Patients with DFUs were enrolled in an IRB-approved study and followed for 30 days. MSI wound images were captured during the initial patient visit. Healing assessments were then obtained for each DFU at the end of 30 days of standard wound care therapy. DFUs were considered healed if the percent-area-reduction was at least 50% at 30 days. A deep learning algorithm, convolutional neural network, was trained using MSI data for input and 30-day healing assessments as the target to predict DFU healing. Performance of the trained model was evaluated using cross-validation. A total of 29 images of DFUs were obtained and used for algorithm development. Of these, 16 DFUs (55%) failed to heal after 30 days. The trained algorithm was able to predict non-healing DFUs with 75% sensitivity and 75% specificity. This is significantly better than the reported clinician accuracy of approximately 50%. Algorithm improvements are ongoing and include expanding available data for input and investigation of optical markers correlated with DFU healing. This work could lead to improved healing prediction models, decreased stasis periods, and increased healing rates.