Development and validation of a model using thoracic vertebrae data from low-dose CT scans for detecting AI-QCT-defined lumbar osteoporosis

Despite evidence supporting opportunistic lumbar osteoporosis screening with low-dose computed tomography (LDCT), a clinically deployable screening strategy is lacking. This study aimed to develop and validate a model using thoracic bone mineral density (BMD) from LDCT for opportunistic screening. This multi-center, cross-sectional study included 18,298 adults who underwent LDCT with coverage from T1 to L2 at 134 Chinese health check-up centers. Outcome was defined as "artificial intelligence quantitative computed tomography (AI-QCT) defined lumbar osteoporosis" based on a mean L1-L2 BMD < 80 mg/cm³ by phantom-less AI-QCT system. Five machine learning models were trained on T1-T12 BMD (n = 11,065) and evaluated internally (n = 4743) and externally (n = 2490). Performance metrics included accuracy, area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value, negative predictive value, Brier score, F1 score and area under the precision-recall curve (AUPRC). AI-QCT-defined lumbar osteoporosis was present in 1730 (9.5%) participants. Higher thoracic BMD was inversely associated with the AI-QCT-defined lumbar osteoporosis. T12 exhibited the strongest association: each 1 mg/cm³ increment reduced odds by 21.9% (OR: 0.781, 95% confidence interval [CI]: 0.769, 0.793, P < 0.001). Logistic regression was selected as the final model. In external validation, it achieved an AUC of 0.993 (95% CI: 0.991, 0.996), AUPRC of 0.922 (95% CI: 0.887, 0.951), F1 score (0.813), sensitivity (0.975), and specificity (0.963). AI-QCT-derived thoracic BMD showed potential for identifying AI-QCT-defined lumbar osteoporosis. Integrating phantomless AI-QCT into lung cancer screening may enable large-scale opportunistic screening, but prospective validation against independent standards is required.