Physiological Parameter Estimation from Reduced-Wavelength Hyperspectral Data: A Model-Based Study

Abstract

Hyperspectral imaging (HSI) enables non-invasive estimation of physiological indices, including tissue oxygen saturation (StO₂), near-infrared perfusion index (NIR), tissue hemoglobin index (THI), and tissue water index (TWI). However, dense spectral acquisition increases hardware complexity and computational cost. This study investigates whether accurate physiological parameter estimation is feasible using a reduced wavelength subset. A stable 20-band configuration was derived from region of interest (ROI) level analysis of the HeiPorSPECTRAL dataset (5758 samples, 11 subjects). Multiple regression models including Ridge, Support Vector Regression, Partial Least Squares, XGBoost, multilayer perceptron (MLP), and one-dimensional convolutional neural network were evaluated under strict subject-disjoint cross-validation to predict physiological indices. The MLP provided consistently low errors across all physiological indices, with mean absolute errors of 0.0607 (StO₂), 0.0269 (NIR), 0.0437 (TWI), and 0.0313 (THI). Selected wavelengths aligned with known hemoglobin and water absorption regions, supporting physiological plausibility. Pixel-wise reconstruction further demonstrated spatial generalization of ROI-trained models. These findings indicate that reduced-wavelength hyperspectral configurations can approximate full-spectrum physiological estimation by using regression models, supporting the development of compact multispectral systems for clinical use.