Generalizable framework for multi-site bone density prediction using non-dominant wrist optical biomarkers.

Osteoporosis is a common skeletal disorder characterized by reduced bone mineral density (BMD) and elevated fracture risk, particularly in aging populations. This study presents a machine learning framework for intelligent optical bone densitometry (iOBD) using near-infrared (NIR) wrist imaging as a non-invasive alternative to dual-energy X-ray absorptiometry (DXA). Among six imaging configurations (three wavelengths × bilateral wrists), 940 nm left wrist images were selected based on superior tissue penetration and reduced inter-subject variability. Features extracted from preprocessed NIR images using a pre-trained VGG16 network were combined with physiological data to predict BMD at five DXA sites. Model benchmarking across eight feature-regressor combinations identified the VGG16 + KNN model as optimal, achieving a test mean absolute percentage error (MAPE) of 8.5% and a correlation coefficient of r = 0.718 at the one-third distal radius. Prediction performance declined at distant sites, with spine (r = 0.460) and femur (r < 0.15) showing limited generalization, particularly in elderly subjects. These results suggest that iOBD can serve as a radiation-free triage tool in primary care, helping identify who should be referred for confirmatory DXA while lowering cost and scan time.