Skin Tone in Hyperspectral Imaging and Its Implications for Fairness in AI

Abstract

Artificial intelligence (AI) is increasingly applied in healthcare, but concerns remain about bias affecting under-represented groups. We investigated whether skin tone is systematically encoded in hyperspectral imaging data and how this affects classifications. Images were collected from 45 healthy women of the upper leg skin and vulvar mucosal tissue. Skin tones were grouped using the individual typology angle scale. Physiological parameters (oxygen saturation, haemoglobin, water and near-infrared indices) were compared across groups. Unsupervised and supervised classification models were evaluated. Skin tone values ranged from −0.7 to 75.8 (20 very light, 9 light, 9 intermediate, 7 tan and 2 brown). All physiological parameters differed significantly across groups (p < 0.001). Unsupervised learning achieved 38.5% balanced accuracy, whereas supervised learning reached 71.4%, with high accuracies for tan (94.6%) and brown (95.0%) groups. Skin tone influences HSI data; it may act as a confounder in AI models, underscoring the need for diverse datasets to ensure equitable performance.