Big Data/Machine Learning/AI
Development of Self-Assessment Tools for Osteoporosis among Postmenopausal Vietnamese Women: A Machine Learning Approach Thuy Trang Nguyen* Tunglam Nguyen Christine Pallota My Hanh Bui Khuong Quynh Long
Development of Self-Assessment Tools for Osteoporosis among Postmenopausal Vietnamese Women: A Machine Learning Approach
Authors: TungLam Nguyen1, Christine Pallotta2, My Hanh Bui2, Thuy Trang Nguyen2, Quynh Long Khuong3
1Pleasantville High school, NY, USA
2Hanoi Medical University Hospital, Hanoi, Vietnam
3Department of Biostatistics, Epidemiology, and Informatics, Perelman School of Medicine, University of Pennsylvania, PA, USA
Background
Osteoporosis is a major health concern in Vietnam due to a rise in aging rates. However, cost-effective early screening tools tailored to the Vietnamese population are lacking. In this study, we applied a machine learning approach to develop self-assessment tools for osteoporosis in menopausal women.
Methods
We used retrospective data from 16,516 postmenopausal Vietnamese women extracted from electronic medical records. Bone mineral density (BMD) measurements of the lumbar vertebrae (L1–L4) and the left and right femoral necks were obtained using the dual-energy X-ray absorptiometry (DXA) system. Osteoporosis was defined as a BMD T-score of < -2.5. Eight algorithms (Logistic Regression, Decision Trees, Random Forest, XGBoost, K-Nearest Neighbors, Neural Networks, Lasso and Ridge regressions, and Naïve Bayes) were utilized to develop prediction algorithms for each anatomical site. The main predictors included age, menopause age, weight, height, lifestyles, and comorbidities. Cross-validation was employed to prevent overfitting.
Results
The prevalence of osteoporosis, as determined by BMD, varied across anatomical sites, ranging from 37% to 51% in the lumbar vertebrae and from 19% to 21% in the left and right femoral necks, respectively. The performance of the models varied slightly across algorithms, though the differences were not substantial. Considering the balance between performance and simplicity, logistic regression was selected as the final algorithm. The final models were developed using four predictors: age, menopause age, height, and BMI, with the area under the curve (AUC) ranging from 0.75 at lumbar vertebrae L4 to 0.84 at right femur. Our models demonstrated superior performance compared to existing tools, such as OSTA, which were developed for the general Asian population.
Conclusion
The newly developed self-assessment tools were shown to be simple and effective in predicting osteoporosis among postmenopausal Vietnamese women. These tools have the potential to serve as early screening instruments for osteoporosis, particularly in resource-limited settings.