Machine learning reconstruction of digit bone Raman spectra enables noninvasive transcutaneous detection of systemic osteoporosis.

Osteoporosis, a major global epidemic, often remains undetected until a fracture occurs, partly because access to screening with reference standard methods such as dual-energy X-ray absorptiometry (DXA) remains limited. As a potential complementary, nonionizing screening approach, we present transcutaneous spatially offset Raman spectroscopy (SORS) combined with machine learning (ML) to recover bone Raman spectra through overlying soft tissue and extract diagnostically relevant information. In this human cadaveric study, we acquired paired Raman measurements from fingers of normal, osteopenic, and osteoporotic donors at transcutaneous SORS and then from the same exposed finger bones. Using this paired dataset, supervised ML models were trained to reconstruct the exposed bone Raman spectra from transcutaneous measurements. Across 38 cadaveric hands, the reconstructed bone spectra achieved a mean correlation coefficient of r = 0.987 with corresponding exposed bone spectra. The ML-predicted spectra also preserved the characteristic bone Raman features and demonstrated statistically significant bone health-associated differences for key Raman derived metrics, including carbonate substitution and mineral-to-matrix ratios. While the distal-radius DXA T-scores expectedly correlated with the exposed bone spectra (r = 0.88, RMSECV = 0.90), they also correlated with the ML-predicted spectra (r = 0.69, RMSECV = 1.40). To demonstrate translational feasibility, preliminary in vivo transcutaneous SORS measurements on the proximal phalanges of two volunteers revealed clear bone-related spectral features consistent with cadaveric measurements. Together, these results establish a foundation for transcutaneous Raman assessment of systemic bone health using ML spectral reconstruction from accessible finger measurement sites.