Explainable machine learning for infrared spectral analysis of bone

Abstract Infrared spectroscopy provides chemically specific information on bone composition, but conventional analysis often reduces spectra to predefined peak-ratio parameters. Here, we evaluated whether multivariate full-spectrum analysis improves bone phenotype classification. Optical photothermal infrared (O-PTIR) spectroscopy was used as the experimental IR modality, with wild-type (WT) and osteogenesis imperfecta mouse (OIM) bones representing typical and fragile bone phenotypes, respectively. Conventional parameters included mineral content, carbonate content, HPO4 content, mineral crystallinity, and collagen integrity. Classification using a combination of these parameters achieved 59–63% accuracy across partial least squares discriminant analysis (PLS-DA) and machine-learning classifiers, whereas models using the 1800–800 cm-1 fingerprint region reached 87–90%. Models based on the collagen-rich region performed similarly to the full spectrum and better than the mineral-rich region. Similar performance across PLS-DA, logistic regression, and XGBoost showed that classification gains were driven primarily by spectral representation rather than model complexity. Variable importance in projection (VIP) and spectral-zone-based SHapley Additive exPlanations (SHAP) identified prominent contributions from collagen-associated Amide I and II regions, with broader patterns of importance in PLS-DA and logistic regression and more localized attribution in XGBoost. Together, these findings support interpretable multivariate full-spectrum analysis as a complementary approach to conventional peak-ratio analysis of bone IR spectra.