Osteoporosis, often referred to as the “silent disease,” is a common disease that reduces bone density and increases the risk of fractures. Remarkably, there are no obvious symptoms until a fracture occurs, highlighting the urgency of early detection and prevention strategies.¹⁹. Osteoporotic fractures, especially hip fractures, often result in significant morbidity, increased mortality, and significant medical costs.²⁰. The social and economic impact of osteoporosis-related fractures makes it imperative to predict this disease not only from a clinical perspective but also from a public health and economic perspective.^{twenty one}.

Early prediction and identification of osteoporosis can pave the way for timely intervention, potentially slowing or reversing bone loss. This not only reduces the risk of fractures, but also improves the quality of life for older people, ensuring greater independence and lower healthcare expenditures.^{twenty two}. When osteoporosis is identified at an early stage, interventions ranging from lifestyle modifications to drug therapy have been shown to be quite effective.^{twenty three}.

The current study demonstrates the potential of machine learning in advancing osteoporosis prediction and highlights a new approach that combines the power of different predictive algorithms.^{twenty four}. Existing methodologies mainly rely on bone mineral density (BMD) testing using DXA scans.^{twenty five}. Although effective, these tests are not universally accessible, can be cost-prohibitive, and are often performed at the onset of clinical symptoms, potentially delaying timely intervention. there is.

In many practical scenarios, particularly in resource-limited settings such as communities, obtaining comprehensive lifestyle data, laboratory tests, or advanced imaging results can be difficult or prohibitive. . Therefore, building predictive models using data that can be extracted from primary medical records and community surveys provides a promising approach for early screening and detection of osteoporosis in these settings.Chen Li²⁶ Researchers were able to predict the risk of rotator cuff tears in hospital outpatients using simple questionnaire data and physical examination findings using machine learning techniques. Similarly, Limin Wang et al.²⁷ We used health questionnaire metrics and regression algorithms to accurately predict symptomatic knee osteoarthritis. By identifying high-risk patients using simple indicators and recommending more detailed medical examinations, we reduce unnecessary medical tests and contribute to savings in medical costs.

By leveraging Germany’s national primary healthcare data, this study provides a non-invasive and efficient means of predicting osteoporosis risk based on health indicators and chronic diseases. The wide inclusion of patients across different health backgrounds ensures the generalizability of the model and its applicability to real-world settings. Our aim is to develop preclinical models that can contribute to early warning, early detection, and diagnosis of high-risk populations. This study did not include medical test indicators and omics data as predictors. Although this may reduce model performance, it also has the benefit of reducing model complexity and increasing practicality. Innovation in the field of machine learning doesn’t necessarily mean using cutting-edge algorithms or complex feature engineering.²⁸. In some cases, simplifying model development to increase versatility and ease of use can be an important form of innovation. Simple models are easier to replicate and validate for other researchers, and are more realistic to implement in real-world settings. Our research results show that the AUC of the model we developed is 0.76, indicating good predictive performance.

The choice of algorithm in this study was critical to ensuring robust predictive performance. The preliminary selection included various algorithms, among which LR, ADA, and GBC emerged as the top candidates in terms of his AUC metric. Previous studies in medical diagnostics have emphasized the importance of AUC as an indicator of a model’s ability to distinguish between positive and negative cases.²⁹. The study by Meng, Y. et al.³⁰ Our results suggest that sequential models such as GRU and LSTM may perform better than non-sequential models such as LR and XGB, but the benefits of these models are limited in the context of cross-sectional data only. It may not be fully utilized. In this study, considering the characteristics of the dataset used for modeling, we adopted a non-sequential model as the final prediction model, which also achieved good prediction performance. Our findings are consistent with recent literature suggesting the promise of these algorithms in health-related predictive tasks.^{31, 32, 33}.

Ensemble methods have consistently demonstrated the ability to improve predictive accuracy by combining the strengths of multiple models and improving the limitations of individual models.³⁴. Using a stacked ensemble (a model that synergizes the robustness of LR, ADA, and GBC) in our study significantly increased his AUC during internal validation. This approach takes advantage of the clear decision boundaries provided by each algorithm and provides a holistic and comprehensive prediction. This approach provides higher predictive accuracy than individual models. This is a result consistent with contemporary research on ensemble methods.³⁵.

The optimal threshold probability of 0.52 derived from the Youden index emphasizes the balanced consideration of both sensitivity (true positive rate) and specificity (true negative rate) in the study. This not only allows accurate identification of actual osteoporosis cases, but also minimizes false alarms, which is important in clinical applications to avoid overdiagnosis and unnecessary interventions. The lift value of 1.9 achieved with the Stacker model highlights its ability to effectively identify osteoporosis cases compared to random selection and validates its clinical utility.

Additionally, a comprehensive feature selection process and rigorous validation reaffirm the model’s robustness and reliability. Her application of SHAP values ​​to feature importance not only promotes transparency in machine learning predictions, but also provides clinical insights, helping healthcare professionals understand and prioritize risk factors. Helpful.³⁶.

SHAP values, an advanced tool for model interpretability, helped determine the salience of each feature within the predictive framework. This finding resonates with the broader osteoporosis literature, with age and gender emerging as the most important factors.There is a long-established relationship between aging and loss of bone density, making age a highly important predictor of osteoporosis risk.³⁷. Gender differences, particularly postmenopausal changes in women, exacerbate osteoporosis risk, and our model highlights their importance.³⁸.

The importance of impaired lipid metabolism in predicting osteoporosis in our model provided interesting insights. Recent studies have begun to identify a potential association between dyslipidemia and changes in bone mineral density (BMD).^39,40. Lipids are involved in bone metabolism, and abnormal lipid profiles can negatively impact bone health. Our model’s emphasis on cancer as a risk factor highlights the multifaceted relationship between cancer and osteoporosis. Some cancer treatments, especially those that involve hormone therapy, can accelerate bone loss and make patients more susceptible to osteoporosis.^41,42. COPD is also thought to be associated with decreased bone density and increased risk of fractures. Pulmonary dysfunction and decreased BMD share an underlying inflammatory pathway. The chronic inflammatory state of COPD disrupts bone metabolism and may lead to increased risk of osteoporosis⁴³. Hypertension is associated with an increased risk of osteoporosis, which may be due to changes in calcium homeostasis and the effects of antihypertensive drugs.⁴⁴. Stroke patients also face an increased risk of osteoporosis and fractures, likely due to immobilization and nerve damage that affect bone metabolism. Similarly, heart failure, CHD, and chronic kidney disease are all associated with increased risk of osteoporosis and fractures.^{45, 46, 47}.

It has never been clearer that early prediction of osteoporosis is essential. As the global population ages, the public health burden of osteoporotic fractures is likely to increase. Against this background, our study stands as a meaningful advance toward strengthening osteoporosis prediction methods. Leveraging IMS HEALTH’s open-source primary healthcare dataset, which includes a large number of patient records, we sought to develop a machine learning-based predictive model. With further research and validation, we hope that this model will help community health workers to screen patients at high risk of osteoporosis using simple indicators during health follow-up. These high-risk patients are provided with individualized health advice and further medical tests, such as laboratory tests and radiology, are recommended to clarify the diagnosis. This could help reduce unnecessary medical tests and save healthcare costs, while ensuring benefits for osteoporosis patients.

Several algorithms were evaluated in our work, and a stacked ensemble approach combining Logistic Regression (LR), Ada Boost Classifier (ADA), and Gradient Boosted Classifier (GBC) was found to be particularly promising. The dominance of this ensemble model highlights the inherent complexity of osteoporosis prediction. We highlight that the multifaceted nature of this disease may be best captured by leveraging the strengths of multiple algorithms.