Development of Machine Learning Models for Prediction of Osteoporosis from Clinical Health Examination Data

Abstract

Osteoporosis is treatable but often overlooked in clinical practice. We aimed to construct prediction models with machine learning algorithms to serve as screening tools for osteoporosis in adults over fifty years old. Additionally, we also compared the performance of newly developed models with traditional prediction models. Data were acquired from community-dwelling participants enrolled in health checkup programs at a medical center in Taiwan. A total of 3053 men and 2929 women were included. Models were constructed for men and women separately with artificial neural network (ANN), support vector machine (SVM), random forest (RF), k-nearest neighbor (KNN), and logistic regression (LoR) to predict the presence of osteoporosis. Area under receiver operating characteristic curve (AUROC) was used to compare the performance of the models. We achieved AUROC of 0.837, 0.840, 0.843, 0.821, 0.827 in men, and 0.781, 0.807, 0.811, 0.767, 0.772 in women, for ANN, SVM, RF, KNN, and LoR models, respectively. The ANN, SVM, RF, and LoR models in men, and the ANN, SVM, and RF models in women performed significantly better than the traditional Osteoporosis Self-Assessment Tool for Asians (OSTA) model. We have demonstrated that machine learning algorithms improve the performance of screening for osteoporosis. By incorporating the models in clinical practice, patients could potentially benefit from earlier diagnosis and treatment of osteoporosis.

Predicting the risk of osteoporosis in older Vietnamese women using machine learning approaches

“…Ho-Pham et al applied four machine learning models—artificial neural networks (ANN), LoR, SVMs and k-nearest neighborhood—to the BMD hip data of Australian women to identify hip fractures 33 . Ou Yang et al implemented five ML models—ANN, SVM, RF, K-nearest neighbors (KNN), LoR—with many features, which were categorized into different areas related to bone health 34 . This study examined 16 input features for men and 19 input features for women in order to identify the relationship between the presence of certain features and risk of osteoporosis in a Taiwanese population.…”

Artificial intelligence used to diagnose osteoporosis from risk factors in clinical data and proposing sports protocols

“…In 2021, Wang et al 34 studied 1559 Chinese women over the age of 20 to develop an ANN model using age and weight as input for prediction of osteoporosis, achieving an AUROC of 0.78. A study in 2021 by Yang et al 3 entitled Development of Machine Learning Models for Prediction of Osteoporosis from Clinical Health Examination Data used a sample population of 3053 Taiwanese men and 2929 women; in this research, the best AUROC of 0.843 and 0.811 in men and in women, respectively was achieved with the RF algorithm for predicting osteoporosis. The following input characteristics were used: medical history of diabetes and hypertension, history of smoking and alcohol consumption liver function, thyroid function, lipid profile, blood protein content, electrolytes, hematological profile, renal function, and for women, history of obstetrics and gynecology were also included.…”

“…The following input characteristics were used: medical history of diabetes and hypertension, history of smoking and alcohol consumption liver function, thyroid function, lipid profile, blood protein content, electrolytes, hematological profile, renal function, and for women, history of obstetrics and gynecology were also included. Among male patients, secondary causes such as alcohol abuse, steroid therapy, and other metabolic disorders account for up to 65% of cases of osteoporosis 3 . In contrast, the prevalence of secondary osteoporosis is much lower in women than in men 35 .…”

“…In comparison with previous research, this had the benefits of predicting osteopenia and osteoporosis with a larger sample size selection (1224 patients) and the inclusion of more input features (17 features in men and 19 features in women) from different aspects. Reducing the number of variables (features) leads to poorer performance and the inclusion of more features contributes to better performance of AI models 3 . Input features were selected based on known association with bone health and easy availability.…”

“…OP is considered a great public health problem and the most common metabolic bone disease as it causes more than 8.9 million fractures per year, resulting in one fracture every three seconds and affecting more than 200 million people worldwide 2 . To increase physicians' awareness of asymptomatic osteoporosis and to identify at-risk patients, it is crucial to understand the risk factors and appropriately diagnose the disease 3 . Several factors such as gender, age, body mass index (BMI), height, low body weight, adequate levels of physical activity, poor nutritional status, family history, calcium, and vitamin D intake, back pain, and other endocrine and cardiometabolic factors are associated with osteoporosis and very important in diagnosing it during lifetime 4 , 5 .…”