Longitudinal pre-diagnostic blood samples predict osteoporosis risk.

Cross-sectional biomarker discovery studies for degenerative diseases are limited by selection bias, post-diagnostic confounding, and high biological variability. Repeated pre-diagnostic sampling can address these limitations. We analyzed 538 plasma metabolomic profiles (1-6 samples per participant, yearly intervals) from 99 pre-diagnostic osteoporosis cases and 99 matched controls in the Danish Blood Donor Study. We compared performance of cross-sectional and longitudinal modeling using least absolute shrinkage and selection operator (LASSO) logistic regression and random forest. Longitudinal LASSO modeling of metabolite trajectories achieved area under the curve (AUC) values between 0.67 and 0.76, outperforming cross-sectional models with AUC values of 0.57-0.63. Batch- and participant-independent validation of the best model demonstrated stable methodology (AUC: 0.70). The longitudinal data revealed that hippuric acid decreased over time in cases while increasing in controls. We conclude that longitudinal sampling costs more per participant, but the substantial information gain makes it a viable strategy when patient recruitment, not sampling capacity, is the bottleneck.