In-Silico Bone Biopsy: Fusing Machine Learning with a Differentiable Bone Metabolism Simulator

Overview Fragility fractures caused by osteoporosis are a major obstacle to healthy life expectancy in an ageing society. Bone strength is governed not only by bone density but also by bone quality — the structural and material properties of trabecular bone. The clinical standard, DXA, is non invasive but only yields a 2D density measure; obtaining detailed bone quality indices or the underlying metabolic dynamics requires an invasive bone biopsy. Diagnosis therefore faces a hard trade off between non invasiveness and information content. My graduation thesis (Kyoto University, Faculty of Engineering, February 2026; supervised by Prof. Taiji Adachi) proposes In Silico Bone Biopsy (ISBB) : a computational system that estimates and visualizes a patient's internal bone state from non invasively obtainable clinical data. A patent application has been filed on this work. Approach: a modular gray box system Black box machine learning adapts well to individual data but is hard to interpret; white box mathematical models are physiologically consistent but hard to personalize. ISBB loosely couples the two: 1. Machine learning module. A model predicts bone mineral density (BMD) from serum biomarkers, anthropometrics, and demographics, trained on two cycles (1999–2002) of the NHANES cohort (8,722 subjects). Among the candidates, a heteroscedastic regression model — which predicts the mean and the variance — generalized best, so the boundary conditions carry a quantified uncertainty rather than a single point estimate. Predicted BMD and bone turnover markers (uNTx, BAP) are mapped to…