Osteoporosis is diagnosed by DXA scan — a specialized X-ray that measures bone mineral density. The machine costs six figures, lives in a radiology department, and emits ionizing radiation. Most people with osteoporosis don't know they have it until they fracture something. The disease is treatable but underdiagnosed because the diagnostic is inaccessible.
The alternative: shine a laser through a finger. Raman spectroscopy detects molecular vibrations — the chemical bonds in bone produce a characteristic spectrum distinct from the overlying tissue. The challenge is that the finger has skin, fat, and tendon layered over the bone. The bone signal is buried.
Spatially offset Raman spectroscopy addresses this by collecting light at different distances from the illumination point. Light that travels deeper into tissue emerges farther from the laser spot. By measuring at multiple offsets, you get spectra weighted toward different depths. But the bone spectrum is still convolved with tissue contributions, and the deconvolution is messy.
Machine learning replaces the deconvolution. Trained on paired measurements — transcutaneous Raman spectra collected through intact fingers, and ground-truth bone spectra from the same samples — the model learns to reconstruct what the bone spectrum would look like without the tissue in the way. The reconstructed spectra distinguish normal bone from osteoporotic bone (p < 0.05) and correlate with DXA T-scores (r = 0.73).
The significance isn't the accuracy (a correlation of 0.73 is good but not diagnostic-grade). It's the modality: no ionizing radiation, no specialized facility, a measurement that takes seconds on a finger. Preliminary in vivo testing suggests it translates from cadaver samples to living patients. If it works, osteoporosis screening moves from radiology to primary care — from a test you schedule to a test you receive.