A machine learning model using commonly available health information distinguished osteoporosis from osteopenia among Asian adults with low bone mass, according to a study published in Frontiers in Endocrinology.  

Dual-energy x-ray absorptiometry (DXA) provides a definitive diagnosis of osteoporosis, but access remains limited across much of Asia, particularly in rural areas. Researchers therefore sought an alternative screening approach based on clinical data collected during routine care.  

"By utilizing only routine clinical variables, the model addresses the recognized limitation of screening accessibility in resource-limited healthcare settings," investigators wrote. "This approach may serve as a preliminary screening tool to identify high-risk individuals within primary care populations for further definitive testing." 

The retrospective diagnostic study included 1,203 adults with DXA-confirmed low bone mass who underwent health examinations from January 2019 through December 2024 at Shenzhen People’s Hospital in Guangdong, China. Of these, 989 had osteopenia and 214 had osteoporosis.  

The researchers compared 11 machine learning algorithms to identify the approach that best balanced discrimination, stability, and clinical interpretability. The algorithms included logistic regression, random forest, support vector machine with recursive feature elimination, extreme gradient boosting, gradient boosting decision tree, decision tree, multilayer perceptron, linear discriminant analysis, adaptive boosting, Gaussian naive Bayes, and LightGBM.  

Linear discriminant analysis (LDA) provided the most balanced and stable performance among the models evaluated. It achieved an area under the receiver operating characteristic curve (AUROC) of 0.738 during model development and 0.710 in the holdout validation set. The limited decrease in discrimination supported internal stability, although external validation was not performed.  

The investigators also evaluated LDA under three class-imbalance strategies: the original data, class weighting, and the synthetic minority oversampling technique (SMOTE). With the original data, sensitivity was 0.977, but specificity was only 0.108, indicating that the classification threshold strongly favored identification of osteoporosis at the expense of correctly identifying osteopenia. Synthetic minority oversampling increased specificity to 0.609 and reduced accuracy from 0.837 to 0.720, while the AUROC remained stable at 0.712 vs 0.710.  

To interpret the LDA model, the researchers used Shapley Additive Explanation (SHAP) analysis to estimate the contribution of nine predictors in the final model: body weight, age, waist-to-height ratio, serum uric acid, alkaline phosphatase, systolic blood pressure, total cholesterol, mean corpuscular hemoglobin, and height.  

Body weight had the greatest influence on model output, followed by age and waist-to-height ratio. Older age and higher waist-to-height ratio, alkaline phosphatase, systolic blood pressure, and total cholesterol were generally associated with higher model-estimated probability of osteoporosis, whereas greater body weight and higher serum uric acid were associated with lower estimated probability. 

Performance also differed substantially by sex, with the LDA model showing a validation AUROC of 0.766 among women compared with 0.613 among men. 

In a separate comparison with the Osteoporosis Self-assessment Tool for Asians, LDA had a higher AUROC—0.736 vs 0.691. LDA accuracy was 0.724 vs 0.669, and specificity was 0.740 vs 0.674. Sensitivity was similar at 0.650 vs 0.645. 

To support potential clinical use, the researchers converted the model into a nomogram incorporating the nine predictors. The nomogram assigns points to each variable and translates the total into an estimated probability of osteoporosis. Higher scores were associated with lower body weight, higher waist-to-height ratio, lower serum uric acid, higher systolic blood pressure, higher mean corpuscular hemoglobin, greater height, higher total cholesterol, older age, and higher alkaline phosphatase. Calibration showed close agreement between predicted and observed outcomes, and decision-curve analysis indicated a positive net benefit across a wide range of threshold probabilities.  

The study has several limitations. It was retrospective and conducted at a single center, limiting the ability to generalize the findings to other populations. Because the analysis was cross-sectional, the model distinguished existing osteoporosis from osteopenia but could not determine disease progression or predict subsequent fracture outcomes. 

In addition, the investigators did not systematically compare recursive feature elimination with other feature-selection methods. They also acknowledged that multicollinearity among height, weight, and waist-to-height ratio may have confounded some associations, including the counterintuitive positive association between height and osteoporosis. Performance was substantially weaker in men, among whom the prevalence of osteoporosis was 13.9%, contributing to greater class imbalance. 

The researchers emphasized that prospective validation and implementation studies are needed. They also called for multicenter external validation, evaluation of additional routinely accessible biomarkers, and development of sex-specific models to improve diagnostic accuracy. 

The authors reported no commercial or financial conflicts of interest.