Longitudinal Opportunistic T12 CT-BMD Assessment from Serial Health-Checkup Chest CT in Midlife Women

The study shows that AI-derived T11 and T12 BMD values can be automatically extracted from routine health-checkup chest CT series and assembled into individual longitudinal BMD trajectories for each woman. Prior opportunistic BMD assessment has largely relied on cross-sectional measurement from a single CT; this work links repeated CT examinations of the same person into a time series, making longitudinal change an analyzable object. Single-center retrospective cohort of 1,034 women and 4,185 examinations over 3.22 +/- 0.30 years, with slopes estimated by linear mixed-effects models.

The relationship between low baseline bone status and subsequent rate of BMD loss ran opposite to intuition: the low-status group showed smaller absolute loss and a less negative T12 slope. This suggests that a low baseline BMD value cannot be simply extrapolated to faster future loss, challenging the intuitive assumption that a single low measurement predicts rapid subsequent decline. Group differences were statistically significant (10.82 vs 12.52 mg/cm3, P = 0.022; -3.08 vs -3.91 mg/cm3/year, P = 0.011), supported by T11-only and protocol-consistent sensitivity analyses.

The between-group difference in relative (percentage) change was not robust, failing to persist after covariate adjustment or when using T11 only. This indicates that relative-change metrics are sensitive to baseline dependence, the definition of low baseline status, and covariate adjustment, so absolute and relative change can lead to different conclusions. Based on comparison between unadjusted analysis and adjusted plus T11-only sensitivity analyses within the same dataset, functioning as a robustness check.

Adding T11 BMD to clinical variables provided no clear incremental discrimination for identifying rapid relative loss (>= 3%/year). This indicates that, in its current form, the AI-derived CT-BMD measure has not yet demonstrated predictive gain beyond routine clinical variables. Prediction modeling was exploratory; AUC rose from 0.708 to 0.716 with P = 0.128, not reaching statistical significance.

The results apply to women aged 45-65 who underwent health checkups at a single center and had multiple noncontrast chest CT examinations; their value lies in demonstrating the feasibility of building individual BMD trajectories from existing imaging, and in providing a starting point for later testing of clinical utility in settings that have dual-energy X-ray absorptiometry or phantom-calibrated quantitative CT reference standards together with repeatability assessment and external validation.

Readers should still note that the 128 mg/cm3 low-bone-status threshold is an exploratory CT-derived classification rather than a clinical diagnosis of osteoporosis or osteopenia; the absence of reference-standard comparison and repeatability assessment leaves the measurement error range of AI-derived CT-BMD unclear; the relative-change conclusion is sensitive to baseline dependence and covariate adjustment, suggesting that different modeling choices could change the directional interpretation; and both the rapid-loss threshold and the prediction model are exploratory, with performance in external populations and on other scanners yet to be verified. The available text is summary-level and does not include figures or supplementary material, so stratified results and model details could not be further checked.