Illuminate-OxyView: a handheld point-of-care device for imaging tissue oxygen saturation in diabetic foot ulcers using a skin-tone-specific machine learning approach

Chronic wounds such as diabetic foot ulcers (DFUs) affect almost 20 million people worldwide often resulting in limb amputation rates of around 20% and 5-year mortality rates of 50-70%. Depending on the pathophysiology of the ulcer, DFUs can take beyond 3 months to heal. A steady supply of oxygen to the tissues surrounding the wound is crucial for wound healing. However, in DFUs, the supply of oxygen is compromised in 50-70% of the patients due to diabetes-related complications such as peripheral vascular disease. Therefore, to effectively track DFU healing, it is critical to have a rapid, point-of-care method to spatially image the tissue oxygen saturation (StO₂). Here, we present Illuminate-OxyView, an AIpowered rapid, point-of-care handheld imager that houses NIR-reflectance-based StO₂ imaging to assess DFUs in low resource settings. Oxygenated and deoxygenated hemoglobin (HbO₂ and Hb) exhibit significant differences in their near infrared (NIR) absorption spectra at certain wavelength bands. This results in a difference in NIR reflectance which is used to infer the spatial distribution of StO₂ in and around the wound. However, reflectance based StO₂ inference is often confounded by melanin concentration in the outermost layer of the skin which varies across different population groups. Owing to this, it is crucial to have a skin-tone specific imaging technique. Therefore, we capture multispectral reflectance at wavelengths 𝜆 = 660 nm, 740 nm, and 850 nm and, in addition, a white light reflectance for the prediction model to infer the skin tone. We conducted a clinical study spanning 300 patients where white light images were used to first detect the Fitz-Patrick score of the skin-tone and, subsequently, NIR reflectance characteristics were processed by a Random- Forest regressor to determine the StO₂ at each point. We demonstrate a >90% correlation of the predicted StO₂ with gold standard FDA-approved StO₂ imager with a root mean squared error (RMSE) of less than 10% StO₂. In summary, we demonstrate an accurate Fitz-Patrick score-specific handheld point-of-care Machine learning (ML)-based StO₂ imager for DFU management in low resource settings.