Osteoporosis is a systemic skeletal disease, and as one of the most prevalent metabolic disorders, its pathogenesis is characterized by a decrease in the amount of bone per unit volume, which leads to fractures. Osteoporosis has, therefore, received progressively increased attention in orthopedics and endocrinology (1, 2). In recent years, the prevalence of osteoporosis has risen due to population aging and the extension of average human lifetime (3). Each year, ~75 million people worldwide have osteoporosis (4). Projections show that by 2050, Asia is expected to have the highest prevalence of osteoporosis, accounting for 50% of global osteoporotic fractures (5). It is worth pointing out that the management of osteoporosis places a heavy burden on the economy. In the United States, the estimated cost of this burden ranges between $13.7 and $20.3 billion; in the European Union, the cost is as high as €31 billion; and in the Asia-Pacific region, the figure reaches $13 billion (6–8). The condition results in diminished mobility and wellbeing, along with the potential for fragility, fractures, and mortality. This disorder adversely affects the lives of elderly individuals with osteoporosis and imposes a considerable medical and economic cost on society. Osteoporosis is frequently undetected in both the early and late phases of the illness, resulting in many patients receiving a diagnosis only after experiencing a fragility fracture (9, 10).

Confirmation of the diagnosis of osteoporosis and assessment of bone mineral density (BMD) involves a variety of technological tools, such as ultrasonography, dual-energy CT (DECT), dual-energy X-ray absorptiometry (DXA), and multichannel convolutional neural network (MCNN)-based processing of raw radiofrequency signals from quantitative ultrasound (QUS) (11–14). Among these methods, DXA has been used as the technique of choice for assessing BMD and calculating T-scores for the diagnosis of osteoporosis due to its high accuracy and wide acceptance (15, 16). According to the guidelines issued by the World Health Organization (WHO) in 1994, osteoporosis can be diagnosed in postmenopausal women and men over 50 years of age if their T-score is not higher than −2.5 standard deviations (T ≤ -2.5 SD) (17). However, the critical technology for DXA testing is X-rays, which leads to limitations that potentially affect multiple systems of the body and a wide range of diseases, problems that cannot be circumvented in current treatments (18–20). Given this, early osteoporosis screening, prediction, and diagnosis are particularly critical. There is an extremely urgent need to develop scientific, rational, and easy-to-use tools for early clinical screening, prediction, and diagnosis.

Shim et al. (21) used machine learning models using gradient boosting machine (GBM), support vector machine (SVM), artificial neural network (ANN), and logistic regression (LR) methods for osteoporosis risk prediction model development, and the optimal model, ANN, was derived with a model sensitivity of 0.741 under five-fold cross-validation. Lee et al. (22) used GridSearchCV or RandomizedSearchCV to measure each model's AUC, accuracy, and F1 scores using five-fold cross-validation after selecting the optimal hyper-parameter combinations for models such as LR. A prediction model for osteoporosis based on the LR algorithm was finally constructed with a model accuracy of 0.75. While these models exhibit reasonable performance, their predictive ability may be limited by the model selection and feature inclusion strategies employed. To compensate for these shortcomings, our study introduces a novel ensemble model using a combinatorial algorithm. The diagnostic logic of the model is also made more relevant to social life by innovatively incorporating social factors (e.g., education level).

The clinical data used in this study were derived from the paper “Construction and Validation of a Nomogram Clinical Prediction Model for Predicting Osteoporosis in an Asymptomatic Elderly Population in Beijing.” In the study, the researchers constructed a clinical nomogram prediction model for osteoporosis using SPSS 26.0 and R 4.0.2 software, which was designed to assist clinicians in quickly recognizing whether a patient has osteoporosis. The study results showed that the three parameters of gender, education level, and body weight have significant predictive value for the diagnosis of osteoporosis, which can assist physicians in making a rapid and effective diagnosis. This study used machine learning techniques to retest the original study by applying 10 machine learning methods and combining them to form 135 machine learning models to establish the optimal clinical prediction model. In addition, by applying the SHAP technique, we further elucidated the importance of each parameter in the model. This approach differs from the underlying nomogram model used in earlier studies in that we utilize a combination of machine learning algorithms to achieve a higher level of prediction, along with the deployment of the Shiny program to enable online disease prediction. These methodological innovations were not present in earlier studies and represent a significant improvement in prediction performance and clinical applicability.

Ultimately, this study developed a set of diagnostic tools for osteoporosis based on the Shiny platform. It aims to improve clinicians' diagnostic accuracy and alleviate the potential side effects and financial burden associated with dual-energy X-ray absorptiometry (DXA) testing.

Unlike conventional statistical analysis approaches in medicine, machine learning techniques forecast new observations by obtaining knowledge from existing information. Nevertheless, numerous sophisticated machine learning models exhibit considerable requirements for transparency and interpretability. To clarify the predictive and evaluative mechanisms behind machine learning models, explainable artificial intelligence (XAI) techniques have been utilized in clinical research. Among these, the SHAP (Shapley Additive exPlanations) method quantifies the extent and direction of variable contributions to the predicted outcomes of machine learning models, offering a visual representation of these contributions (23).

This study utilized patients' clinical markers to predict osteoporosis through a machine learning (ML) model, with the model's prognostic results clarified by SHAP technology. This study's principal findings and contributions are summarized below:

The insidious nature of osteoporosis poses a challenge for early intervention, especially in the elderly patient population, where traditional DXA testing methods are not only invasive but also costly due to their specific physiological and physical conditions, which limits widespread screening for the disease in developing countries and regions. Therefore, how to effectively predict whether an elderly patient has osteoporosis quickly and accurately at an early stage in order to guide clinical personalized treatment has been a significant focus and difficulty in medical research.

The rise of quality and personalized medicine has placed artificial intelligence-driven prediction models in the vanguard of clinical research. This study employed a retrospective analysis to gather health data on senior participants and assess the prevalence of osteoporosis among them. We developed and validated a clinical prediction model for osteoporosis in asymptomatic individuals using machine learning techniques. The therapeutic efficacy of the created model was rigorously evaluated using contrast analysis and internal confirmation, aiming to enhance early detection and personalized treatment for elderly osteoporosis patients, hence assisting physicians in making more informed decisions. And compared to the traditional multicollinearity test, this study used the elastic network regularization and SHAP techniques. Elastic network regularization effectively reduces the effect of covariance by applying penalties to redundant features, while the SHAP technique evaluates the importance of features and filters out key predictors, thus reducing the risk of multicollinearity. In addition, 10-fold cross-validation is used to ensure model robustness and prevent overfitting.

This research employed 10 machine learning techniques to develop 135 predictive models. The integrated machine learning methods of GBM+RF, GBM+SVM, SVM+FDA, NN+FDA, NN+SVM, NN+RF, KNN+SVM, KNN+RF, Bayes+SVM, and SVM+CART shown strong performance in diagnosing osteoporosis patients. Through a thorough evaluation of the predictive efficacy of various models, we determined that the KNN+RF model exhibits superior classification performance and robustness in detecting osteoporosis (OP), achieving the highest AUC value of 90.4%. Other recent studies have proposed relevant prediction models; for example, Jang et al. (25) developed a DNN-based deep learning model using imaging and clinical data with an AUC of 0.867. Similarly, Carvalho et al. employed a comprehensive machine learning model incorporating a large number of biochemical metrics, achieving an AUC of 0.94 (26). In contrast, our model utilizes only non-invasive features and achieves an AUC of 0.904, highlighting the practical efficiency and robustness of our model, which does not rely on imaging or biochemical data. This indicates enhanced predictive capability regarding the presence or absence of osteoporosis, leading us to select the KNN+RF model for further analysis. The calibration plots indicated that the predicted curves of the combined KNN+RF model corresponded with the observed curves. DCA plots indicated that employing KNN+RF, SVM+CART, and the integrated NN+RF model for the diagnosis and prediction of OP, along with suitable therapeutic actions, is advantageous for patients in clinical practice. The impact of the identified characteristics in the KNN+RF model on OP was elucidated using the SHAP technique, revealing that the four metrics of Weight, Waistline, Height, and Education level exerted the most substantial influence on the diagnosis of OP. This finding is consistent with previous studies (27, 28).

Relevant literature has found that for osteoporotic fractures, low body weight is one of its causative risk factors, especially for older menopausal women, and wasting is one of the main factors leading to osteoporosis (29–31). This explains the higher percentage of female patients (72.4% female) in the statistical analysis of this study. Furthermore, studies have shown that body fat distribution at different sites is positively correlated with bone density, regardless of the site (femur or lumbar spine) (32). The possible reason for this is that different types of fat (e.g., android fat and gynoid fat) are involved in the endocrine regulation of bone benefits (33). Also, low body weight may lead to a decrease in body muscle mass, which can induce sarcopenia and increase the risk of falls, injuries, and fractures, leading to the complication of osteoporosis (34), which is consistent with the results of the baseline table analysis and SHAP analysis in the present study. Moreover, the protective effect of weight gain against osteoporosis can be attributed in part to muscle-derived actin, such as irisin, which has been shown to promote osteoblast differentiation and bone formation, and is positively correlated with bone mineral density (35, 36). Additionally, adipose tissue serves as an endocrine organ, secreting hormones such as adiponectin. In several studies, adiponectin has been shown to be inversely correlated with bone mineral density and adiponectin is also inversely correlated with fat content, so that weight gain will lead to an increase in bone mineral density from the point of view of endocrine regulation, which will in turn become a protective factor against osteoporosis (37, 38). These mechanisms support the observations made in our model. Gkastaris and Zhang et al. demonstrated that obesity significantly impacts society and is strongly associated with osteoporosis. The obesity risk factor is waist circumference (WC), one of the most critical risk factors for the development of osteoporosis. Among them, waist circumference (WC) is an important indicator used to assess the accumulation of abdominal fat, which is associated with the onset of many diseases (9, 39–43). In the present study, Waistline was negatively associated with the prevalence of osteoporosis, which is consistent with the findings of Murat and Saşak (44). This phenomenon differs from conventional wisdom, but it has been suggested that there is an obesity paradox in some populations, whereby moderately overweight individuals may have a better prognosis for osteoporosis, especially in women (45).

Ono et al. (46) and Pouresmaeili et al. (47) have shown that height is an independent risk factor for elderly patients with osteoporosis and that changes in height are a common clinical manifestation in elderly patients with osteoporosis accompanied by vertebral fractures and kyphosis. In addition, the level of education is also associated with the risk of developing osteoporosis. This study showed that individuals with higher levels of education have a better economic status, are more health-conscious, and are therefore less likely to develop osteoporosis. Wang et al. (48) showed a higher prevalence of osteoporosis in older adults with lower education. Meanwhile related studies point out that this may be because less educated people tend to have poorer knowledge about osteoporosis prevention and are more likely to adopt unhealthy lifestyles, including less robust health literacy (unwillingness to take medication), poorer preventive behaviors (e.g., insufficient calcium intake, lack of time and access to physical activity, etc.), and poor diets, among other conditions. In developing countries, the prevalence of osteoporosis is significantly higher than in developed countries. This difference may reflect differences in urbanization, socioeconomic status (SES), healthcare, and health education, with scholars such as Du demonstrating that lower SES and education levels are associated with a higher risk of osteoporosis (49, 50).

We have designed a state-of-the-art Shiny application to diagnose the presence of osteoporosis in the elderly population in order to facilitate its application and dissemination in real-world clinical practice. The application is a clinical decision support tool that provides user-friendly outputs, including personalized predictive probabilities and risk stratification, designed to help clinicians triage patients or guide follow-up testing. Clinicians can enter readily available patient data and obtain immediate predictions to support triage or follow-up recommendations. In the Shiny online prediction model, a prediction probability >0.5 is used as the threshold for clinical intervention. This reflects the default binary classification decision boundary, allowing for straightforward interpretation: patients with model-estimated probabilities >0.5 are considered to be at higher risk for osteoporosis, and further diagnostic evaluation or prophylactic treatment is recommended to clinicians. As shown in Figures 6A, B, the app is based on seven authoritative diagnostic guidelines to assist clinicians in calculating and assessing the individualized risk of developing osteoporosis. The application can be accessed via the following link: https://osteoporosispredictionmodel.shinyapps.io/medic_predict/. The shinyapps.oi server is running as shown in Supplementary Figure S1.

In clinical practice, models should be constructed and applied with holistic considerations rather than relying solely on a single feature for diagnostic prediction. Therefore, the involvement of a panel of experts is essential to assist in the diagnosis of whether a tester has osteoporosis. In addition, our findings are highly consistent with medical literature published worldwide, further validating the model's clinical relevance.

This research possesses certain limitations. First, the generalizability of the model in this study is limited by the small sample size, the fact that the data were collected from only some of the communities in Beijing, and the broad definitions of certain variables (e.g., smoking and alcohol consumption). In the future, external validation should be combined with multi-center and larger-scale data, and model calibration methods should be introduced to improve stability. Second, the model did not incorporate important clinical factors, such as comorbidities and biochemical indicators (e.g., vitamin D and serum calcium), and was based on only four anthropometric indicators, which may not fully reflect the complex etiology of osteoporosis. Third, FRAX is a risk assessment tool recommended by international guidelines (e.g., NOF, IOF) and contains clinical factors and optional BMD data (51). In contrast, the model in this study relies only on basic signs and demographic characteristics, which are suitable for resource-limited scenarios. Although not directly compared with FRAX, the predictive performance of both should be evaluated in the future, and their integration potential should be explored to enhance clinical utility. Weight, Waistline, Height, and Education level are independent predictors of osteoporosis in elderly patients. The clinical prediction model constructed in this study based on these four independent predictors can realize the accurate diagnosis of osteoporosis patients and assists physicians in devising a more evidence-based treatment plan to enhance patient prognosis and minimize societal health expenditures. Although our model predicts good results, some overfitting may occur due to issues such as data limitations. It should be used as a broad screening tool, and actual diagnosis still requires expert input and incorporation of other relevant clinical evidence. Future studies will include prospective validation using DXA as the gold standard, as well as collection of clinician feedback to assess usability, acceptance, and consistency with diagnostic outcomes. Moreover, in future studies, we will further delve into the correlations among the predictors to enhance the model's ability to identify and explain variable interactions.

This study used 10 machine learning methods such as Enet, LR, CART, RF, SVM, Bayes, KNN, NN, FDA, and GBM. We constructed 135 different machine learning models by combining them in order to realize the diagnosis of elderly osteoporosis patients. This research examines the efficacy of machine learning methodologies in clinical forecasting. The study's results indicate that machine learning methods perform effectively in diagnosing osteoporosis, with the combined KNN and RF model exhibiting the most superior classification efficacy and robustness, while all other model metrics also demonstrate commendable performance. Furthermore, we have created a Shiny-based online application for osteoporosis diagnosis, designed to aid clinicians in devising a more logical treatment strategy, minimizing the adverse effects linked to DXA testing technology, thus lowering healthcare expenses and enhancing patient outcomes. For future studies we will incorporate multicenter datasets to further validate the robustness and generalizability of the model.