Machine learning-assisted osteoporosis detection using radiographic canal bone ratio (CBR-7) and demographic features.
Authors
Affiliations (2)
Affiliations (2)
- Department of Orthopaedic Surgery, Tokai University School of Medicine, Kanagawa, Japan. [email protected].
- Department of Orthopaedic Surgery, Tokai University School of Medicine, Kanagawa, Japan.
Abstract
Osteoporosis remains underdiagnosed in aging populations due to its asymptomatic nature and limited access to diagnostic imaging. The Canal Bone Ratio measured 7 cm distal to the lesser trochanter (CBR-7), obtainable from standard full-length spinal radiographs, has emerged as a potential surrogate marker for bone quality. This study investigates the utility of CBR-7 in combination with machine learning (ML) models for opportunistic osteoporosis screening. We retrospectively analyzed 162 patients (mean age, 73.3 years) who underwent DXA and standing full-length spinal radiography before lumbar spine surgery. Osteoporosis was defined as the lower of the lumbar spine and femoral neck T-scores being ≤ -2.5. Six regression models were evaluated using nested five-fold cross-validation to estimate the CBR-7 value corresponding to this diagnostic threshold. Separately, a logistic regression model incorporating CBR-7, age, sex, and BMI was evaluated using stratified five-fold cross-validation to classify DXA-defined osteoporosis. Of the 162 patients, 63 (38.9%) met the DXA-based definition of osteoporosis. CBR-7 was significantly higher in the osteoporosis group than in the non-osteoporosis group (0.517 ± 0.062 vs. 0.468 ± 0.058, p < 0.001). In the threshold-estimation analysis, Linear Regression showed the lowest cross-validated prediction error (MSE = 0.0033 ± 0.0006; R<sup>2</sup> = 0.181 ± 0.117) and estimated a CBR-7 value of 0.499 ± 0.001 corresponding to a Low T-score of -2.5. The logistic regression model combining CBR-7, age, sex, and BMI achieved an out-of-fold AUC of 0.778. At the Youden-derived probability threshold of 0.317, sensitivity was 82.5%, specificity was 57.6%, accuracy was 67.3%, precision was 55.3%, and the F1 score was 0.662. CBR-7 was associated with DXA-defined osteoporosis and may serve as a preliminary radiographic marker for identifying patients who warrant further assessment with DXA. The simple linear model provided a stable estimate of the CBR-7 value corresponding to the diagnostic T-score threshold, whereas the classification model demonstrated moderate discrimination. External validation in independent cohorts, together with prospective multicenter validation, is required before clinical implementation.