Machine learning-enhanced wavelength detection for point-of-care optical devices in tissue oxygenation and peripheral arterial disease assessment

Peripheral Arterial Disease (PAD), a widespread vascular condition affecting millions in the United States, remains underdiagnosed despite its significant healthcare burden. Tissue oxygen saturation (StO₂), assessed through near-infrared (NIR) spectroscopy, offers a promising diagnostic biomarker, but its effectiveness is challenged by confounding factors such as melanin concentration and epidermal thickness variability. To address these limitations, this study developed a 3D MCMATLAB-based in silico model of the human forearm, simulating light-tissue interactions using Monte Carlo methods. By varying parameters such as melanin content, epidermis thickness, and wavelength, the model provided insights into the effects of these variables on photon fluence at a clinically relevant depth. <br/> <br/> Machine learning techniques, including recursive feature elimination, were applied to identify optimal NIR wavelengths for reliable PAD diagnosis. Logistic regression with L1 regularization demonstrated robust feature selection, balancing accuracy and generalizability. The standardized dataset highlighted the potential of combining fluence values across selected wavelengths to enhance classification accuracy. Results showed that despite the different melanin concentrations and epidermis thicknesses that were used in the study, the models were able to detect StO₂ with an R² of 0.892 and a mean absolute error of 1.205 and were able to classify the PAD and without PAD classes with an accuracy of 66.11&percnt;. This interdisciplinary approach highlights a pathway toward developing accessible, non-invasive diagnostic tools for diverse populations.