Development and evaluation of a machine learning model for osteoporosis risk prediction in Korean women

Abstract

Background

The aim of this study was to develop a machine learning (ML) model for classifying osteoporosis in Korean women based on a large-scale population cohort study. This study also aimed to assess ML model performance compared with traditional osteoporosis screening tools. Furthermore, this study aimed to examine the factors influencing the risk of osteoporosis through variable importance.

Methods

Data was collected from 4199 women aged 40–69 years in the baseline survey of the Ansan and Ansung cohort of the Korean Genome and Epidemiology Study. Osteoporosis was set as the dependent variable to develop ML classification models. Independent variables included 122 factors related to osteoporosis risk, such as socio-demographic characteristics, anthropometric parameters, lifestyle factors, reproductive factors, nutrient intakes, diet quality indices, medical history, medication history, family history, biochemical parameters, and genetic factors. The six classification models were developed using ML techniques, including decision tree, random forest, multilayer perceptron, support vector machine, light gradient boosting machine, and extreme gradient boosting (XGBoost). The six ML classification models were compared with two traditional osteoporosis screening tools, including the osteoporosis risk assessment instrument (ORAI) and the osteoporosis self-assessment tool (OST). The ML model performances were evaluated and compared using the confusion matrix and area under the curve (AUC) metrics. Variable importance was assessed using the XGBoost technique to investigate osteoporosis risk factors.

Results

The XGBoost model showed the highest performance out of the six ML classification models, with an accuracy of 0.705, precision of 0.664, recall of 0.830, and F1 score of 0.738. Moreover, the XGBoost model showed a higher performance on AUC than ORAI and OST. Variable importance scores were identified for 69 out of the 122 variables associated with osteoporosis risk factors. Age at menopause ranked first in variable importance. Variables of arthritis, physical activities, hypertension, education level, income level; alcohol intake, potassium intake, homeostatic model assessment for insulin resistance; energy intake, vitamin C intake, gout; and dietary inflammatory index ranked in the top 20 out of the 69 variables, using the XGBoost technique.

Conclusions

This study found that an XGBoost model can be utilized to classify osteoporosis in Korean women. Age at menopause is a significant factor in osteoporosis risk, followed by arthritis, physical activities, hypertension, and education level.

Background

Osteoporosis is a skeletal disease characterized by systemic disorders of bone mass content and microstructure [1]. The reduction of bone mass and density in osteoporosis patients increases the risk of osteoporotic fractures that lead to high mortality [2]. It is vital to prevent osteoporosis to alleviate the social and economic burden due to osteoporotic fractures [3].

The risk of osteoporosis is rising with the increasing elderly population [4]. A recent meta-analysis of 108 studies performed across six continents showed that global prevalence of osteoporosis subjects was estimated at 19.7% [5]. The global prevalence of osteoporosis subjects aged over 50 years was approximately 2.3 times higher in women (26.0%) than in men (11.2%) [5]. According to data from the Health Insurance Review and Assessment Service, the prevalence of osteoporosis subjects in Korea was estimated at approximately 1.047 million in 2020 and 94.3% were women [6].

According to the World Health Organization (WHO), the diagnostic criterion for osteoporosis is defined as bone mineral density (BMD) at the lumbar spine or hip that is 2.5 standard deviations or more below the average BMD of healthy young adults [7]. The BMD is measured using radiology methods, including dual X-ray absorptiometry (DXA), quantitative ultrasound (QUS), and quantitative computed tomography (QCT) [8].

There are simple screening tools to help predict the risk of osteoporosis. The osteoporosis risk assessment instrument (ORAI) evaluates osteoporosis risk by considering age, weight, and previous use of hormone replacement therapy [9]. The osteoporosis self-assessment tool (OST) uses weight and age as key indicators to assess the prediction of osteoporosis risk [10]. However, osteoporosis is influenced by various risk factors beyond age, weight, and previous use of hormone replacement therapy, and simple osteoporosis screening tools have limitations including low sensitivity and specificity [2, 11].

A decrease in bone mass is caused by an imbalance in bone remodeling through various factors [12]. Osteoporosis is related to various factors such as age and gender [13], genetics [14], medical and medication histories [15], and nutrient intake [16]. In women, menopause plays an important role in a bone mass decrease. Estrogen decreases the rate of bone remodeling activation and helps maintain the stability between bone formation and resorption [17]. However, a decreased level of estrogen in postmenopausal women can increase osteoclast activity and accelerate bone loss by 3–5% per year over 5 to 10 years, thereby increasing the risk of osteoporosis [17,18,19].

Machine learning (ML), a branch of artificial intelligence, is a computerized process that can classify and predict data patterns through learning experience from data [20]. Several recent studies were conducted to predict risks of hypertension [21, 22], dyslipidemia [23], type 2 diabetes mellitus [24], and breast cancer [25] using ML techniques. A few studies using ML techniques conducted for osteoporosis prediction [26,27,28]. These studies were performed using osteoporosis risk factors such as age, anthropometric and biochemical parameters. Inui et al. [26] developed a ML model to predict osteoporosis in 2541 elderly women without DXA data using 24 variables such as body mass index (BMI) and 22 biochemical parameters. Bui et al. [27] developed an osteoporosis prediction ML model in 1951 elderly Vietnamese women with 15 variables including height, weight, 11 biochemical parameters, and geographical location. Ou Yang et al. [28] conducted a ML model to predict osteoporosis in 5982 elderly Taiwanese, with 16 variables for men and 19 variables for women including height, weight, waist circumference, history of alcohol consumption, history of smoking, 2 medical histories, 3 obstetrics and gynecology history (for women), and 8 biochemical parameters.

Studies on osteoporosis prediction using ML techniques have also been conducted in Korean women [29, 30]. Kwon et al. [29] developed a ML model to predict osteoporosis in 1431 postmenopausal Korean women utilizing Korea National Health and Nutrition Examination Survey (KNHANES) data conducted in the national, cross-sectional study. They included age, education level, 5 anthropometric parameters, 6 biochemical parameters, 4 lifestyle factors, and 3 reproductive factors. Similarly, a study by Shim et al. [30] employed an osteoporosis prediction ML model in 1792 postmenopausal Korean women using the KNHANES data. They included osteoporosis risk factors, such as age, 4 anthropometric parameters, 4 lifestyle factors, 3 reproductive factors, and 7 medical histories.

Despite these prior efforts, few studies have extensively classified and predicted osteoporosis using ML techniques in Korean women, including various factors such as socio-demographic characteristics, anthropometric parameters, diet quality indices, nutrient intakes, reproductive factors, lifestyle factors, family history, medical history, medication history, biochemical parameters, and genetics factors based on a large-scale population cohort study.

Therefore, the aim of this study was to develop a ML model to classify osteoporosis using multiple variables related to osteoporosis in Korean women based on a large-scale population cohort study. This study also aimed to evaluate the performance of the ML models in comparison with traditional osteoporosis screening tools. Moreover, this study aimed to examine the importance of variables to clarify to what extent factors influence the risk for osteoporosis.

Methods

Study population

This study utilized baseline survey data (2001 to 2002) from the Ansan and Ansung cohort study of the Korean Genome and Epidemiology Study (KoGES) conducted by the National Institutes of Health at the Korea Disease Control and Prevention Agency [31]. The KoGES is a large-scale cohort study that collects various data on socio-demographic characteristics, anthropometric parameters, genetic factors, lifestyle factors, dietary assessment, biochemical parameters, medical history, medication history, family history, and reproductive factors and performs follow-up studies [31].

The Ansan and Ansung cohort study comprises a baseline survey conducted from 2001 to 2002 and an 8th follow-up survey. The baseline data of the Ansan and Ansung study, part of the population-based cohorts in the KoGES, included 10,030 men and women and performed biennial surveys of residents aged 40 to 69 years in Ansan (urban) and Ansung (rural) [31, 32].

We included women aged 40 to 69 years from the baseline data of the Ansan and Ansung study in KoGES, which involved 10,030 subjects. We excluded subjects with missing data on energy intake (n = 781), men (n = 4451), missing data on menopause status (n = 28) and missing data on single nucleotide polymorphism (SNP) (n = 571). Finally, 4199 subjects were included (Fig. 1).

Fig. 1

A flow-diagram of the study subjects. KoGES, Korean Genome and Epidemiology Study; SNP, single nucleotide polymorphism

This study was conducted following the guidelines of the Declaration of Helsinki, with all subjects providing written informed consent. This study was approved by the Institutional Review Board of Gyeongsang National University (GIRB-A22-NX-0073) and the Korean Health and Genomic study at the Korea National Institute of Health (NBK-2023-003).

Independent variables

Table 1 presents 122 independent variables in the 11 categories. We divided 11 categories associated with osteoporosis risk factors, such as socio-demographic characteristics, anthropometric parameters, lifestyle factors, nutrient intakes, diet quality indices, medical history, medication history, family history, reproductive factors, biochemical parameters, and genetic factors.

Table 1 Independent variables of this study

We deemed the disease occurred if the subjects responded “yes” to the following questions: “Have you been diagnosed with the disease by a doctor?” or “Have you been currently treated for the disease?” Based on studies showing the association between medical history, medication use, and osteoporosis [15, 33], we included 14 medical history variables, such as hypertension, diabetes mellitus, allergic diseases, myocardial infarction, thyroid disease, congestive heart failure, coronary artery disease, hyperlipidemia, asthma, chronic obstructive pulmonary disease, kidney disease, cerebrovascular disease, gout, and arthritis (osteoarthritis and rheumatoid arthritis). We considered subjects as having taken the medications if they answered “yes” to the following questions: “Have you been taking medication continuously?” or “Have you experienced taking medication?” We also included 13 medication history variables such as steroids, oral contraceptives, hormone replacement therapy, anticonvulsants, anticoagulants, and medications of insulin, hypertension, arthritis, thyroid, osteoporosis, stroke, asthma, and hyperlipidemia. The family history of osteoporosis was divided into parents, siblings, others, and none. Anthropometric parameters included measuring subjects’ height (cm) to the nearest 0.1 cm and body weight (kg) to the nearest 0.1 kg, with subjects wearing light clothing without shoes. The BMI was calculated as weight (kg) divided by height squared (m²). The body fat and muscle mass were assessed by bioelectrical impedance analysis (Inbody 3.0, Biospace, Seoul, Korea). The blood pressure was measured in a sitting position with the arm at heart level in a stable state. The homeostatic model assessment for insulin resistance (HOMA–IR) was calculated using fasting glucose and fasting insulin variables [34]. The estimated glomerular filtration rate (eGFR) was calculated using serum creatinine [35]. Genetic data were obtained through the Affymetrix Genome-wide Human SNP Array 5.0.

Under the categories of nutrient intakes and diet quality indices, the semi-quantitative food frequency questionnaires (SQFFQ) comprised 103 food items, assessing the frequency of each item over the past 12 months [31]. The frequency was categorized as follows: never or seldom, once a month, two or three times a month, one or two times a week, three or four times a week, five or six times a week, once a day, twice a day, or three times or more a day. Nutrient intake calculation (energy, carbohydrate, protein, fat, fiber, retinol, beta-carotene, vitamin D, vitamin E, vitamin K, vitamin B6, vitamin C, calcium, sodium, phosphorus, potassium, magnesium, iron, zinc, copper, manganese, selenium, and cholesterol) was performed with the SQFFQ data by computer aided nutritional analysis program (CAN–Pro) 5.0 software (The Korean Nutrition Society, Seoul, Korea).

Diet quality indices were calculated to estimate the impact of dietary patterns on osteoporosis including index of nutritional quality (INQ [36]), net endogenous acid production (NEAP [37]), potential renal acid load (PRAL [37]), alternate Mediterranean diet score (aMED [38, 39]), Dietary Approaches to Stop Hypertension (DASH [40, 41]), dietary inflammatory index (DII [42]), and Korean healthy eating index (KHEI [43]).

Dependent variables

Osteoporosis, the dependent variable, was evaluated based on the T-scores of the distal radius and the midshaft tibia BMDs using the QUS device Omnisense 7000 S/P (Sunlight Medical Ltd, Petah Tikva, Israel). We included 806 subjects with osteoporosis and 3393 subjects without osteoporosis.

Development environment and data preprocessing

The first step was data preprocessing. Missing values were imputed with the mode values for categorical variables and the mean values for continuous variables. We encoded the dataset using the categorical boosting (CatBoost) encoder, which is advantageous for handling large-scale datasets and transforming categorical and string data into continuous scalar data [44].

Osteoporosis ML classification models were implemented in Python (version 3.9.13) using libraries such as NumPy (version 1.25.2), pandas (version 1.5.3), scikit-learn (version 1.2.2), and category-encoders (version 2.6.3).

Model development and evaluation of model performance

We developed an osteoporosis ML classification model using six ML techniques, including decision tree, random forest, multi-layer perceptron (MLP), support vector machine (SVM), light gradient boosting machine (LGBM), and extreme gradient boosting (XGBoost).

Among these techniques, we primarily employed two advanced gradient boosting models, LGBM and XGBoost, as our main models while incorporating traditional ML algorithms for comparison. The selection of these models was driven by several key considerations. LGBM and XGBoost models were chosen due to their established advantages in handling structured tabular data. These models demonstrated superior performance in processing large-scale datasets with high dimensionality, which aligns well with the characteristics of our dataset. In addition, both models have exhibited robust performance on imbalanced datasets through built-in mechanisms for handling class imbalance. Their leaf-wise tree growth strategies and sophisticated regularization techniques could effectively prevent overfitting while maintaining model performance.

Meanwhile, for comparison, traditional ML algorithms such as decision tree, random forest, MLP, and SVM models were selected. These models provide a comprehensive benchmark against traditional ML methodologies, allowing us to evaluate whether the computational complexity of advanced gradient boosting techniques offers meaningful improvements over conventional approaches. These classical ML models were selected based on their distinct characteristics: decision tree for interpretability, random forest for ensemble robustness, MLP for complex non-linear relationship modeling, and SVM for effectiveness in high-dimensional spaces.

To further enhance the performance of the ML models, we conducted hyperparameter tuning was conducted using the Optuna library for model training and optimal parameter discovery. Optuna [45] is an automated software framework developed for efficient hyperparameter optimization, utilizing the Tree-structured Parzen Estimator (TPE) technique based on Bayesian Optimization. The library defines the search space through user-specified objective functions and evaluates attempted hyperparameter combinations to identify the optimal configuration.

Optuna’s key features include dynamic search space definition, handling of complex constraints, and support for parallelization, enabling efficient and flexible hyperparameter optimization. These capabilities are particularly valuable in maximizing model performance when working with large-scale datasets or complex models [45].

To train and evaluate the ML models, we split the dataset into training and test datasets. The data was imbalanced with 806 subjects having osteoporosis and 3393 subjects without osteoporosis. To address this imbalance, we randomly selected 100 subjects with osteoporosis and 100 subjects without osteoporosis for the testing dataset. The remaining data was used as the training dataset. The traditional osteoporosis screening tools, ORAI and OST, were calculated for evaluation and comparison with six ML models [9, 10].

Table 2 shows the accuracy, precision, recall, and F1 score for evaluating model performance based on the confusion matrix. Accuracy measures the percentage of correctly predicted cases, indicating how well the predictions match the actual outcomes. Precision indicates the percentage of true positive instances out of the cases predicted as positive. Recall represents the proportion of correctly predicted positive instances out of all actual positive instances, reflecting the extent to which true positive outcomes were detected. The F1 score is a performance metric calculated as the harmonic mean of precision and recall.

Table 2 Confusion matrix and performance evaluation metrics

We evaluated and compared the performance of ML classification models and traditional osteoporosis screening tools by utilizing the area under the curve (AUC) metrics. The AUC provides a comprehensive measure of a model’s predictive accuracy by quantifying its ability to distinguish between different classes based on the model’s prediction probabilities.

The AUC serves as a single scalar value that quantifies the model’s overall predictive performance. AUC values range from 0 to 1, with higher values indicating superior classification performance. Specifically, an AUC of 0.5 suggests no discriminative ability (equivalent to random guessing), while values above 0.7 are generally considered indicative of useful predictive capability. Calculation of the AUC involves analyzing the model’s performance across various discrimination thresholds, considering metrics such as true positive rate (TPR, sensitivity) and false positive rate (FPR, 1-specificity).

XGBoost technique for variable importance

We used the XGBoost technique to assess the importance of various factors affecting osteoporosis. Variable importance was evaluated using 122 variables related to osteoporosis risk factors, including socio-demographic characteristics, lifestyle factors, anthropometric parameters, reproductive factors, nutrient intakes, diet quality indices, medical history, medication history, family history, biochemical parameters, and genetic factors.

Statistical analysis

The normality of distribution was assessed using the Kolmogorov-Smirnov test, Q-Q plots, and histograms. Log transformation was applied to variables that did not show normal distribution. Continuous variables were analyzed using two-sample t-test for normally distributed variables, while the Mann-Whitney U test was used for non-normally distributed variables. Categorical variables were evaluated using chi-squared analysis. Normally distributed variables are presented as means ± standard errors, while non-normally distributed variables are shown as medians and interquartile ranges. Statistical analyses were conducted using SPSS 27.0 (IBM, Chicago, IL, USA).

Results

Characteristics of study subjects

The characteristics of the study subjects are presented in Table 3. Out of a total of 4199 women, 806 subjects had osteoporosis. The osteoporosis group (median: 61 years) was older than the non-osteoporosis group (median: 49 years). The osteoporosis group (92.2%) had a higher proportion of postmenopausal women compared with the non-osteoporosis group (56.3%). The non-osteoporosis group showed significantly higher education and income levels than the osteoporosis group. The osteoporosis group had a significantly higher prevalence of hypertension, diabetes mellitus, arthritis (osteoarthritis and rheumatoid arthritis), and gout compared with the non-osteoporosis group. Moreover, the osteoporosis group had a significantly higher level of C-reactive protein (CRP) and HOMA–IR than the non-osteoporosis group. Under the categories of nutrient intakes and diet quality indices of the study subjects, the non-osteoporosis group had significantly higher intakes of energy, protein, and fat than the osteoporosis group. The osteoporosis group had significantly lower intakes of calcium, phosphorus, selenium, retinol, and beta-carotene than the non-osteoporosis group. The non-osteoporosis group consumed more vitamin B6, vitamin D, vitamin E, and vitamin K than the osteoporosis group. In diet quality indices, DII was significantly lower in the non-osteoporosis group compared with the osteoporosis group.

Table 3 Characteristics of the study subjects

Model performance

Table 4 presents the performance comparison results of six ML classification models (decision tree, random forest, MLP, SVM, LGBM, and XGBoost) and two traditional osteoporosis screening tools (ORAI and OST) using ML techniques. The XGBoost model showed the highest accuracy, precision, recall, and F1 score out of the six ML classification models, with an accuracy of 0.705, precision of 0.664, recall of 0.830, and an F1 score of 0.738 (Table 4). Moreover, the XGBoost model showed a higher accuracy, precision, and F1 score than the two traditional osteoporosis screening tools. Figure 2. presents the receiver operating characteristic curves of six ML classification models and traditional osteoporosis screening tools. The six ML classification models showed a higher AUC than traditional osteoporosis screening tools. The XGBoost model had the highest performance, with an AUC of 0.84 (Fig. 2.).

Table 4 The results of classification models using machine learning techniques on validation data

Fig. 2

The Receiver operating characteristic curve of six classification models and two traditional osteoporosis screening tools. LGBM, light gradient boosting machine; MLP, multi-layer perceptron; ORAI, osteoporosis risk assessment instrument; OST, osteoporosis self-assessment tool; SVM, support vector machine; XGBoost, extreme gradient boosting

Variable importance

The variable importance assessed using the XGBoost technique to evaluate factors contributing to osteoporosis risk is presented in Fig. 3. We found that 69 out of the 122 variables showed variable importance scores. Age at menopause ranked first in the variable importance. Arthritis (osteoarthritis and rheumatoid arthritis) ranked second, hypertension ranked 5th, and gout ranked 17th in the variable importance. Five physical activities ranked 3rd, 4th, 7th, 20th, and 29th in the variable importance, respectively. Renin and HOMA–IR ranked relatively high, with variable importance rankings of 11th and 14th, respectively. Education level ranked 6th, while income level and age ranked 8th and 13th in the variable importance, respectively. The variable of siblings in the family history of osteoporosis ranked 31st.

Fig. 3

Variable importance derived from the XGBoost technique. AS1_AGE, age; AS1_ALBUMIN_TR, albumin; AS1_ALT_TR, alanine aminotransferase; AS1_AMED, alternate Mediterranean diet score; AS1_ARRM, arthritis (osteoarthritis and rheumatoid arthritis); AS1_BDCMSC, muscle mass; AS1_BDFTR, body fat; AS1_BETACARO, beta-carotene intake; AS1_BPSIT1DIA, diastolic blood pressure; AS1_BRCA, breast cancer surgery; AS1_BUN_TR, blood urea nitrogen; AS1_CALCIUM, calcium intake; AS1_CARBO, carbohydrate intake; AS1_COPPER, copper intake; AS1_CREATININE_TR1, creatinine; AS1_CRP, C-reactive protein; AS1_DII, dietary inflammatory index; AS1_DRHT, hypertension medication; AS1_DRINK, alcohol intake status; AS1_EDUA, education level; AS1_EGFR, estimated glomerular filtration rate; AS1_ENERGY, energy intake; AS1_FE, iron intake; AS1_FIBER, fiber intake; AS1_FMOSREL_S, family history of osteoporosis (siblings); AS1_GT, gout; AS1_HDL_TR, high-density lipoprotein cholesterol; AS1_HEIGHT, height; AS1_HIP, hip circumference; AS1_HOMAIR, homeostatic model assessment for insulin resistance; AS1_HT, hypertension; AS1_ICOFF_1, frequency of coffee consumption; AS1_INCOME, income level; AS1_INSM, insomnia; AS1_KHEI, Korean healthy eating index; AS1_MAGN, magnesium intake; AS1_MN, manganese intake; AS1_NA, sodium; AS1_PHOSPHO, phosphorus intake; AS1_PHYACTH, high-intensity physical activity; AS1_PHYACTL, low-intensity physical activity; AS1_PHYACTM, moderate-intensity physical activity; AS1_PHYSIT, sedentary physical activity; AS1_PHYSTB, stable physical activity; AS1_PMAG_C, age at menopause; AS1_PMYN_C, menopausal status; AS1_POTASSIUM, potassium intake; AS1_PRAL, potential renal acid load; AS1_PREG, pregnancy experience status; AS1_PROTEIN, protein intake; AS1_RENIN, renin; AS1_RETINOL, retinol intake; AS1_SE, selenium intake; AS1_SODIUM, sodium intake; AS1_TCHL_TR, total cholesterol; AS1_TG_TR, triglyceride; AS1_TOTALC, alcohol intake; AS1_TOTPRT, total protein; AS1_VITC, vitamin C intake; AS1_VITD, vitamin D intake; AS1_VITE, vitamin E intake; AS1_VITK, vitamin K intake; AS1_WAIST, waist circumference; AS1_WBC, white blood cell; AS1_ZN, zinc intake; SNP_A-1,809,518, rs628948; SNP_A-1,850,320, rs12590815; SNP_A-2,130,710, rs238340; SNP_A-4,262,878, rs746219

Nutrient intake variables showed relatively high importance. Potassium intake ranked 12th, while energy, vitamin C, and vitamin D intakes ranked 15th, 16th, and 18th, respectively. Moreover, diet quality indices such as DII, aMED, PRAL, and NEAP were placed at 19th, 32nd, 39th, and 62nd in the variable importance, respectively. The genetic factors of rs746219, rs12590815, rs238340, and rs628948 showed relatively low importance, ranking 41st to 44th, respectively.

Discussion

This study aimed to develop a ML model for osteoporosis classification in Korean women based on a large-scale population cohort study. We also aimed to evaluate ML model performance compared with traditional osteoporosis screening tools, including ORAI and OST. Furthermore, we aimed to investigate factors affecting osteoporosis risk in Korean women by examining variable importance.

We found that XGBoost model had the highest performance out of the 6 ML classification models, including decision tree, random forest, MLP, SVM, LGBM, and XGBoost. Moreover, we found that 6 ML classification models had higher AUC performance than 2 traditional osteoporosis screening tools, including ORAI and OST. The XGBoost model had the highest performance of AUC out of the 6 ML classification models and 2 traditional osteoporosis screening tools.

Several studies developed decision tree, logistic regression, random forest, k-nearest neighbor (KNN), SVM, neural network, artificial neural network (ANN), LGBM, and gradient boosting trees of ML models in osteoporosis risk prediction [26,27,28]. Inui et al. [26] employed 5 ML models including decision tree, logistic regression, random forest, LGBM, and gradient boosting trees. They found that LGBM model had the highest performance out of the five ML models [26]. Bui et al. [27] developed 4 ML models, such as logistic regression, random forest, SVM, and neural networks. They found that random forest model had the highest performance out of the 4 ML models. Oh Yang et al. [28] developed 5 ML models, including ANN, SVM, random forest, KNN, and logistic regression. They found that random forest model had the highest area under the receiver operating characteristic curve (AUROC) out of the 5 ML models [28].

Furthermore, previous studies [29, 30] have developed several ML models, including random forest, decision tree, logistic regression, gradient boosting machine, SVM, ANN, adaptive boosting (AdaBoost), and KNN, to predict osteoporosis risk in Korean women. In 3 ML models of random forest, gradient boosting machine, and AdaBoost for the prediction of osteoporosis risk with postmenopausal Korean women which developed by Kwon et al. [29], the Adaboost model showed the highest performance out of the 3 ML models. Moreover, Shim et al. [30] developed 7 ML models, including random forest, decision tree, logistic regression, gradient boosting machine, SVM, ANN, and KNN for the prediction of osteoporosis risk in postmenopausal Korean women. They found that an ANN model had the highest AUROC value out of the 7 ML models [30].

We found variable importance scores for 69 out of the 122 variables associated with osteoporosis risk factors. In our findings, age at menopause and the subject’s age ranked 1st and 8th in the variable importance, respectively, aligning with previous research demonstrating the crucial role of age in osteoporosis risk. A recent prospective longitudinal study showed that women with early menopause and premature ovarian insufficiency (31.3%) had an approximately 1.43 times higher risk of osteoporosis compared with women with usual age at menopause (21.8%) [46]. In a cross-sectional study with 2224 Chinese women aged 40 to 80 years, the association between earlier menopause and the prevalence of osteoporosis was observed [47]. A reduction in estrogen levels after menopause could cause an imbalance in bone formation and resorption, leading to bone loss and an increased risk of osteoporosis [18, 48].

Socio-demographic factors also play a significant role in osteoporosis risk. In our analysis, education level and income level ranked 6th and 8th in the variable importance, respectively. Consistent with our findings, a recent study showed that a higher education level was significantly associated with a heel BMD increase, while reducing the risk of osteoporosis [49]. Moreover, a higher income level was significantly associated with a femoral neck BMD increase [49]. This indicated that higher education and income levels could benefit bone health by increasing access to healthcare and contributing to healthier lifestyles [49]. Moreover, previous cross-sectional studies have shown the association between higher education level or income level and reduced risk of osteoporosis [50, 51].

In addition to variables of socio-demographic characteristics, medical history variables showed strong associations with osteoporosis risk. Arthritis (osteoarthritis and rheumatoid arthritis) ranked 2nd in the variable importance. Arthritis is classified into osteoarthritis and rheumatoid arthritis. Osteoarthritis is a degenerative condition that asymmetrically affects knee and hip joints, while rheumatoid arthritis is a systemic autoimmune disease that impacts small joints, such as hands and feet [52, 53]. Arthritis was associated with inflammatory cytokines, such as interleukin-1 (IL-1), interleukin-6 (IL-6), and tumor necrosis factor-alpha (TNF-α) [54,55,56]. Increased levels of inflammatory cytokines could lead to elevated bone resorption, increasing the risk of osteoporosis [54,55,56]. Moreover, the autoimmune response of the immune system could damage bone and cartilage, increasing the risk of osteoporosis [55, 56]. In line with our findings, a cross-sectional study with 4311 subjects showed that subjects with moderate to severe osteoarthritis were associated with lower T-scores of the lumbar spine and total hip compared with non-osteoarthritis [57]. Moreover, a cross-sectional study of 1322 Korean postmenopausal women with rheumatoid arthritis showed that 619 (46.8%) subjects were diagnosed with osteoporosis [58]. However, our study did not examine arthritis variables separately into osteoarthritis and rheumatoid arthritis. Future studies are needed to examine separately osteoarthritis and rheumatoid arthritis.

Moreover, other medical conditions were significantly associated with osteoporosis risk. Among them, hypertension emerged as a significant risk factor, with hypertension, hypertension medication, and renin levels ranking 5th, 9th, and 11th, respectively, in the variable importance. Consistent with our findings, a retrospective study showed that hypertension was significantly associated with osteoporosis risk in 2039 Chinese postmenopausal women [59]. One potential mechanism linking hypertension to osteoporosis involves the renin-angiotensin system. Renin, an enzyme secreted by kidneys contributes to the production of angiotensin II, which mediates vasoconstriction, leading to blood pressure elevation. This angiotensin II could interfere with bone formation and reduce BMD, thereby increasing the risk of osteoporosis [60, 61].

Other metabolic disorders have also been linked to osteoporosis risk. We found that gout ranked 17th in the variable importance. In line with our finding, a longitudinal study by Kwon et al. [62] showed that subjects with gout had an 11% increased risk of osteoporosis compared with subjects without gout.

Lifestyle factors were identified as key contributors to osteoporosis risk. Among them, physical activities played a crucial role, ranking 3rd, 4th, 7th, 20th, and 29th, respectively, in the variable importance. Consistent with our findings, a recent cross-sectional study found that moderate-intensity physical activity and high-intensity active physical activity could decrease the osteoporosis risk [63]. Physical activities could enhance BMD and bone strength by stimulating the bones, which could reduce the risk of osteoporosis [64, 65].

Alcohol intake was one of lifestyle factors that showed a significant association with osteoporosis. It ranked 10th in the variable importance. A meta-analysis showed that subjects who daily had 1 to 2 alcoholic drinks were likely to have a 1.34 times higher osteoporosis risk than non-alcohol drinkers [66]. Alcohol intake can inhibit osteoblast formation and stimulate osteoclast activity, which causes a decrease in bone formation and an increase in bone resorption [67, 68]. In addition, excessive alcohol consumption can elevate parathyroid hormone levels and induce oxidative stress, causing bone loss [67, 68].

Dietary intake was identified as an important determinant of osteoporosis risk. Nutrient intake variables ranked in relatively high positions in variable importance. Potassium intake ranked 12th in the variable importance. Moreover, potassium intake ranked 1st in the variable importance out of the 23 nutrient intake variables. A cross-sectional study showed that higher daily potassium intake was significantly associated with a 32% reduced risk of lumbar spine osteoporosis in 5142 postmenopausal women [69]. Dietary potassium could neutralize excess acid produced during metabolic processes, thereby maintaining the body’s acid-base balance and supporting mechanisms that promote bone health [70]. Vitamin C, beta-carotene, and zinc ranked 16th, 21st, and 28th, respectively, in the variable importance. Consistently, a cross-sectional study by Kim et al. [71] found positive associations between the intakes of beta-carotene, zinc, and vitamin C and bone health in 189 postmenopausal Korean women.

Dietary quality played a significant role in osteoporosis risk. We investigated the importance of dietary quality indices for osteoporosis in women. DII ranked 19th in the variable importance, which was consistent with the findings from recent studies [72, 73]. A cross-sectional study by Li et al. [72] showed that a higher DII was significantly associated with BMD loss in the femoral neck, intertrochanter, and total hip compared with low DII. The DII is a tool to assess the effect of an individual diet on the level of inflammation in the body [74, 75]. Consumption of pro-inflammatory foods, such as processed meats, refined grains, and high-fat dairy products, could reduce osteoblast function, activate osteoclast activity, and increase inflammation levels, leading to an increased risk of osteoporosis [74, 75].

Biochemical parameters were significantly associated with osteoporosis. Fourteen variables under the category of biochemical parameters appeared to play important roles in osteoporosis. Fourteen variables were renin, HOMA–IR, total cholesterol, eGFR, blood urea nitrogen, triglycerides, albumin, white blood cell, sodium, total protein, creatinine, CRP, high-density lipoprotein cholesterol, and alanine aminotransferase. HOMA-IR ranked 14th in the variable importance, the second highest importance after renin (11th) under the category of biochemical parameters. A prospective study by Napoli et al. [76] found that BMD increased with higher HOMA–IR in 2398 elderly adults without diabetes. This study found that CRP ranked 59th in the variable importance. A recent study by Little-Letsinger et al. [77] found a weak association between CRP and BMDs in the femoral neck and lumbar spine.

Bone turnover markers (BTMs) further contribute to understanding osteoporosis risk. In patients with bone diseases, BTMs, biomarkers found in blood and/or urine, can be used to examine the bone status [78, 79]. BTMs are classified as bone formation markers and bone resorption markers. Bone formation markers include type 1 procollagen-N-propeptide (P1NP), bone-specific alkaline phosphatase (BSAP), and osteocalcin [78, 79]. Bone resorption markers include C-terminal telopeptide of type 1 collagen (CTX), tartrate-resistant acid phosphatase 5b (TRAP 5b), and N-telopeptide of type 1 collagen (NTX) [78, 79]. In a cross-sectional study of 2327 elderly subjects aged 60 to 85 years, BSAP and NTX were inversely associated with lumbar spine BMD [80]. However, data from the Ansan and Ansung study used in this study did not provide BTMs such as CTX and P1NP.

Genetic factors also played a role in osteoporosis risk. We found 4 SNPs (rs746219, rs12590815, rs238340, and rs628948) out of the 12 SNPs ranked 41th to 44th in the variable importance. A recent study by Park et al. [14] found the association between SNPs and the risk of osteoporosis.

This study has several strengths. We utilized the Ansan and Ansung cohort study from KoGES, which is large-scale general Korean population-based cohort data, in order to construct a ML model for osteoporosis classification. We attempted to include as many osteoporosis risk factors as much as possible in the ML models. These ML models have the potential to be applicable in screening women with a high risk of osteoporosis. These ML models could be used in the field of early detection of osteoporosis, identifying risk factors, and allowing personalized osteoporosis prevention strategies. Therefore, ML models could enhance osteoporosis related-health outcomes in women, which could be beneficial in clinical and community settings.

Despite the strengths, this study has limitations. The Ansan and Ansung study of KoGES used in this study collected data using self-reported questionnaires, which could potentially have recall bias. The osteoporosis ML classification model was developed using baseline survey data from the Ansan and Ansung study of KoGES. Further studies are necessary to validate the ML classification model with follow-up data from the Ansan and Ansung study of KoGES.

Conclusions

In conclusion, we developed a ML model to classify osteoporosis in Korean women using various osteoporosis risk factors. The ML classification model using the XGBoost technique outperformed the ML classification models using the decision tree, random forest, MLP, SVM, and LGBM techniques and traditional osteoporosis screening tools using the ORAI and OST. In the variable importance using the XGBoost technique, age at menopause was the most crucial osteoporosis risk factor.