Predicting Osteoporosis Risk in Older Vietnamese Women Using Machine Learning Approaches

Fragility fractures and their consequences are the most common signs of osteoporosis, the most common skeletal-related disease in adults. Identifying patients at high risk for fracture before a fracture occurs is an essential part of osteoporosis care. This problem continues to be a major concern for researchers and physicians around the world. Despite the fact that several algorithms have been created to identify people with osteoporosis or predict their risk of fracture, concerns remain about their accuracy and usefulness. Scientific breakthroughs, such as machine learning technologies, are rapidly gaining acceptance as alternative approaches to improving risk assessment and existing practices.

Osteoporosis is a serious disease that mainly affects postmenopausal women. The gold standard test for osteoporosis includes estimation of bone mineral density (BMD) in the proximal femur, lumbar spine, and in some cases the forearm using X-ray absorptiometry dual energy (DXA). The BMD is then compared to that of a reference group, including a healthy, premenopausal adult population matched by sex and ethnicity (for example, how much lower it is compared to the standard deviations or to the T score) for the diagnosis1.2. Professional organizations, such as the International Society for Clinical Densitometry, the United States Preventive Services Task Force, and the International Osteoporosis Foundation, all promote screening strategies for older women, but determining when and how to perform screenings is more controversial. .

The Osteoporosis Self-Assessment Tool (OST) is one of the oldest and easiest ways to identify people at risk for osteoporosis. This tool uses ballast and age identify men and women from diverse populations at risk of osteoporosis3,4,5,6,7,8,9,10,11,12,13. For the Asian female population, prediction tools have been developed by integrating the magnitude of the correlation between age and ballast with BMD to estimate the likelihood of osteoporosis5,14,15.

Expanding the number of factors used to determine osteoporosis, 12 input parameters, such as demographics, lifestyle and medical history, have been included in the Fracture Risk Assessment Tool (FRAX)16. Similarly, other complex tools, including ORAI17SCORE18ORISIS19BONE20 and more21, have incorporated additional features to improve the performance of OST detection. Although the combination of many recognized risk factors was expected to enhance the usefulness of screening tools, studies have found that basic tools, such as OST, might work just as well as those with more complicated algorithms, while recent systematic reviews have highlighted the potential and limitations of these tools. approaches.

Machine learning models to predict osteoporosis risk

Scientists around the world recognize that osteoporosis is an important public health problem. Although therapy may reduce fracture risk by 33% to 50%22, only a small percentage of patients, including those with previous osteoporotic fractures, receive proper diagnosis and treatment. In addition, the accurate and timely identification of high-risk and/or high-cost patients should improve effective healthcare management, as well as clinical decision-making and improve service planning and policy.23.24. Over the past few decades, machine learning models have been increasingly integrated into osteoporosis prediction, along with effective uses of big data in healthcare, leading to improved quality and the effectiveness of health care planning and delivery. Additionally, early identification of diseases (through simpler intervention and treatment, for example), personalized health management, and effective detection of fraudulent behavior in healthcare are some of the potential benefits.25. Artificial intelligence (AI) technology has been developed based on mathematical modeling over the years. AI software has been applied to different disciplines including epidemiological investigation26drug discovery27and diagnostic radiology28. At this point, computer-assisted devices have been incorporated into routine clinical practice to detect abnormalities related to respiratory diseases on chest X-ray images.29,30,31. On-site implementation of AI software is proving its benefits in minimizing diagnostic bias, overcoming burnout issues, and improving active case finding in the community. Codlin et al. included 12 AI software to predict tuberculosis on chest X-ray images in the independent performance evaluation, which indicated that half of the AI ​​software had higher specificity values ​​than an intermediate radiologist28.

In a study by Erjiang et al.32, seven machine learning models (CatBoost, eXtreme Gradient Boosting, Neural Networks (NN), Bagged Flexible Discriminant Analysis, Random Forest (RF), Logistic Regression (LoR) and Support Vector Machines (SVM)) were implemented to derive the best fit models to differentiate between patients with and without osteoporosis using DXA T-scores. Ho-Pham et al. applied four machine learning models – artificial neural networks (ANN), LoR, SVM and k-nearest neighborhood – to hip BMD data from Australian women to identify hip fractures33. Or Yang et al. implemented five ML models – ANN, SVM, RF, K-nearest neighbors (KNN), LoR – with many features, which were classified into different areas related to bone health34. This study examined 16 entry characteristics for men and 19 entry characteristics for women to identify the relationship between the presence of certain characteristics and the risk of osteoporosis in a Taiwanese population. Other machine learning methods using OST to predict osteoporosis have been reviewed by Ferizi et al.1.

Osteoporosis and resulting fragility fractures are recognized as major public health problems in many developing countries, especially Vietnam. The lack of DXA equipment to diagnose osteoporosis in these countries necessitates a prediction model for individualized assessment. Ho-Pham et al. proposed a prediction model for individualized assessments of osteoporosis based on age and ballast for men and women14. In this study, a LoR model using data from a population in Ho Chi Minh City was applied to develop the tool for each sex, with good accuracy. The researchers developed and validated a prediction model based on age and ballast to estimate the absolute risk of osteoporosis in the Vietnamese population14. However, few studies have focused on the prediction of osteoporosis based on OST in other regions of Vietnam. The main objective of our study was to build tools to assess the risk of osteoporosis from OST data in women over 50 years old in northern Vietnam. On the other hand, little is known about the performance of the model proposed by Ho-Pham et al.14 when applied to a new population. Ho-Pham’s model has proven its good accuracy during internal validation14 while its performance in predicting osteoporosis in external validation has not been significantly reported. Therefore, our secondary objective was to independently validate the Osteoporosis Self-assessment Tool for Asians (OSTA) model and the model developed by Ho-Pham et al.14 on a new population. In addition to age and ballastour data now includes height, geographic location (urban/rural area) and blood test results of uric acid, cholesterol, creatinine, FT4, glucose, HbA1c, Ure, AST, TSH, calcium and by GGT. In their study, Ou Yang et al. concluded that one specific blood test parameter is relevant for OST (eg, creatinine) in predicting osteoporosis, while others (eg, TSH) were of insignificant value in predicting osteoporosis. osteoporosis in the Taiwanese population34. However, unlike Ou Yang et al. regarding the influence of TSH in North American patients, Jamal et al. recommended that patients suspected of osteoporosis based on their OST score undergo the TSH test34.35. Therefore, the third objective of our research was to validate Ou Yang’s conclusion34 and jamal35 using the data set collected at Hanoi Medical University Hospital as well as to uncover other novel factors related to OST outcomes.

Significance of the study

The results of this study would provide valuable evidence to bolster the potentials of machine learning algorithms as decision support tools in widespread osteoporosis screening. The study would also present the supporting findings to promote the digital transformation of medical diagnosis in Vietnam. Moreover, the covariates that showed the significant contribution to osteoporosis risk would be highlighted and could be a valuable consideration for government policy makers in Vietnam.

Sherry J. Basler