Discovering Vitamin-D-Deficiency-Associated Factors in Korean Adults Using KNHANES Data Based on an Integrated Analysis of Machine Learning and Statistical Techniques

Abstract

Background/Objectives: Vitamin D deficiency (VDD) is a global health concern associated with metabolic disease and immune dysfunction. Despite known risk factors like limited sun exposure, diet, and lifestyle, few studies have explored these factors comprehensively on a large scale. This cross-sectional study aimed to identify VDD-associated factors in South Korea via an integrative approach of machine learning and statistical analyses using Korea National Health and Nutrition Examination Survey (KNHANES) IX-1 data. Methods: Using the KNHANES dataset, six machine learning algorithms were applied to evaluate VDD (serum 25[OH]D3 < 20 ng/mL)-associated factors through feature importance scores. Thereafter, multivariate linear and logistic regression models were applied to the dataset—stratified by sex and age. Results: Among 583 variables, 17 VDD-associated factors were identified using the CatBoost model, which achieved the highest F1 score. When these factors were assessed through statistical analysis, dietary supplement use emerged as a consistent factor associated with VDD across all subgroups (younger men, younger women, older men, and older women). In younger adults, HDL cholesterol, blood and urinary creatinine, water intake, urban residence, and breakfast frequency were significantly associated with VDD. Additionally, blood urea nitrogen and fasting plasma glucose in men and urinary sodium in women showed sex-specific associations with serum 25(OH)D levels. Conclusions: This study identified key VDD-associated factors in the South Korean population, which varied by age or sex. These findings highlight the multifaceted nature of VDD, influenced by dietary, lifestyle, and biochemical factors and underscore the need for strategies integrating machine learning and statistical analysis.

1. Introduction

Vitamin D deficiency (VDD, serum 25-hydroxyvitamin D [25(OH)D] level < 20 ng/mL) is a significant global health issue. For instance, the 2001–2018 National Health and Nutrition Examination Survey (NHANES) revealed that 24.6% of the United States (US) population is affected by VDD [1]. In South Korea, 71.4% of 21,208 participants in the 2010–2014 Korea National Health and Nutrition Examination Survey (KNHANES) had VDD [2]. Furthermore, in India, 83% of healthy adults aged 18–40 years in the Kashmir Valley reportedly had VDD [3]. VDD is well known to increase the risk of rickets and osteomalacia by regulating calcium and phosphorus balance for proper bone mineralization [4]. It has also been linked to metabolic diseases such as diabetes, cardiovascular disease, and cancer, as well as immune-related conditions like infections and autoimmune disorders [5,6,7,8]. Moreover, interest in vitamin D’s role in antiviral immunity has grown following reports of higher COVID-19 incidence in individuals with VDD [9,10].

With growing evidence linking VDD to various chronic diseases, numerous studies have investigated factors associated with VDD worldwide [11,12,13]. For example, an analysis of 55,844 Europeans using the Vitamin D Standardization Program protocol found that 13.0% had VDD, with key factors including geographical differences, sun exposure, and dietary habits [11]. A review of 15 studies involving 2833 pregnant women identified obesity, ethnicity, geographic location, and limited sun exposure as significant contributors to VDD during pregnancy [12]. Similarly, a study of 26,010 U.S. adults using NHANES 2001–2010 data reported strong associations between VDD and non-Hispanic Black race, lower educational attainment, obesity, smoking, physical inactivity, and lower milk consumption [13].

Several studies have also examined VDD risk factors in the South Korean population [2,14,15]. For example, an analysis of Korean adolescents using 2008–2009 KNHANES data identified winter season, older age, higher education level, female sex, obesity, lack of vitamin D supplementation, low milk consumption (0–<200 mL/d), and insufficient physical activity as key risk factors [14]. Another study of 21,208 Korean adults from 2010–2014 KNHANES data found associations between VDD and environmental temperature, ultraviolet radiation, walking frequency, occupation type, and shift work, with age- and sex-specific analyses revealing stronger links between VDD and low walking frequency in younger men and low educational attainment in older women [2]. Moreover, a study of 1594 South Korean nurses of childbearing age reported a high VDD prevalence (89%) and identified age, sampling month, and stress symptoms as significant risk factors [15].

Although multiple risk factors for VDD have been identified, they have not been comprehensively analyzed on a large scale, considering their diversity, ranging from eating and lifestyle habits to biochemical markers. Additionally, the KNHANES has not measured vitamin D status in recent years (2015–2021), leaving the population’s VDD status underreported. Therefore, this study utilized the most recent KNHANES IX data that contains vitamin D status, which was collected in 2022 and released in 2024. Using machine learning techniques, VDD-associated factors were screened among a variety of variables, including biochemical markers, nutrient intake, and lifestyle. Ultimately, the screened factors were statistically assessed to investigate their correlation with VDD using a dataset stratified by age and sex.

2. Materials and Methods

2.1. Data Source and Participants

This study used data from the first round of the 2022 KNHANES IX-1 [16]. The KNHANES, a nationwide cross-sectional study conducted by the Korea Disease Control and Prevention Agency (KDCA), comprehensively assesses the health and nutritional status of the Korean population to inform national policies [17]. The KNHANES IX-1 used a two-stage stratified cluster sampling method, randomly selecting participants from 192 regions across South Korea, with 25 households chosen per region. Of the original 6265 KNHANES IX-1 participants, 943 children aged ≤ 18 years and 171 individuals without blood 25(OH)D3 data were excluded, leaving 5151 participants for the machine learning analysis. To account for potential biological differences, 330 participants without dietary intake data were excluded, and the remaining 4821 were stratified by sex and age as follows: men aged 19–64 years (n = 1424), women aged 19–64 years (n = 1895), men aged ≥ 65 years (n = 671), and women aged ≥ 65 years (n = 831) (Figure 1). The procedures included physical measurements (e.g., height, weight, waist circumference, skeletal muscle mass, and body fat), health behavior surveys (e.g., smoking, alcohol consumption, etc.), laboratory tests using blood and urine samples, and dietary intake data. The KNHANES IX-1 data collection procedure was approved by the Institutional Review Board (IRB) of the KDCA (IRB No. 2018-01-03-4C-A). The IRB of Suwon University waived ethical review and approval requirements for KNHANES IX-1 data analyses (IRB No. 2404-045-01).

2.2. General Characteristics

Body weight and height were measured during a health examination, and body mass index (BMI) was calculated by dividing weight in kilograms by the square of the height in meters. Data on demographic, socioeconomic, and lifestyle factors were obtained via health interviews. Household income was categorized into four quartiles: low, middle-low, middle-high, and high (Q1–Q4). Individuals who consumed alcohol at least once a month in the year preceding the interview were classified as current drinkers. Current smokers were defined as those who had smoked over 100 cigarettes in their lifetime and continued to smoke.

2.3. Laboratory Tests and Dietary Intake Analyses

Blood and urine samples were collected for assorted biochemical analyses. Serum vitamin D levels, including those of 25(OH)D2, 25(OH)D3, and 3-epi-25(OH)D3, were measured using liquid chromatography–mass spectrometry and expressed in ng/mL. Additional blood parameters included fasting blood glucose (mg/dL), total cholesterol (mg/dL), aspartate aminotransferase/serum glutamic–oxaloacetic transaminase (SGOT) (IU/L), hemoglobin (g/dL), blood urea nitrogen (mg/dL), blood creatinine (mg/dL), white blood cell count (Thous/µL), red blood cell count (Mil/µL), platelet count (Thous/µL), uric acid (mg/dL), and high-sensitivity C-reactive protein (mg/L). Urinary measurements encompassed urinary creatinine (mg/dL), urinary sodium (mmol/L), urinary potassium (mmol/L), and urinary albumin (µg/mL).

Dietary intake was assessed using a single 24 h dietary recall. Daily consumption of total energy and nutrients was estimated based on the KNHANES recipe and the food composition database published by the Korean Rural Development Administration [18]. The specific nutrient intake (e.g., vitamin D) used in this study was derived from the processed data provided by KNHANES.

2.4. Machine Learning Analyses

To prepare the machine learning dataset, we excluded variables such as 25(OH)D2, 3-epi-25(OH)D3, and textual responses, resulting in a final dataset (n = 5151) with 583 explanatory variables. The target variable was binary, indicating vitamin D status, classified as deficient or sufficient based on serum 25(OH)D levels. We defined the binary target using the standard cutoff of 20 ng/mL: participants with serum 25(OH)D levels ≥ 20 ng/mL were categorized as sufficient (target = 0), and those with serum 25(OH)D levels < 20 ng/mL were categorized as deficient (target = 1). This threshold aligns with the Institute of Medicine (IOM)’s definition of vitamin D deficiency [19] and is commonly used in numerous previous studies to define vitamin D deficiency and sufficiency [20,21,22,23]. Missing values were imputed via Python’s IterativeImputer from Scikit-learn (v1.3.1).

To classify VDD, six algorithms (Random Forest, GradientBoosting, XGBoost, XGBoost-RF, LightGBM, and CatBoost) were trained and validated using StratifiedShuffleSplit (50 iterations, 20% test set). Specifically, Random Forest, GradientBoosting, and StratifiedShuffleSplit were used from Scikit-learn (v1.3.1); XGBoost and XGBoost-RF from xgboost (v2.0.0); LightGBM from lightgbm (v4.1.0); and CatBoost from catboost (v1.2.2).

The performance of each algorithm was evaluated using the following metrics:

$$Precision={\displaystyle \frac{\mathrm{True} \mathrm{Positives} \left(\mathrm{TP}\right)}{\mathrm{True} \mathrm{Positives} \left(\mathrm{TP}\right)+\mathrm{False} \mathrm{Positives} \left(\mathrm{FP}\right)}}$$

$$Sensitivity={\displaystyle \frac{\mathrm{True} \mathrm{Positives} \left(\mathrm{TP}\right)}{\mathrm{True} \mathrm{Positives} \left(\mathrm{TP}\right)+\mathrm{False} \mathrm{Negative} \left(\mathrm{FN}\right)}}$$

$$Accuracy={\displaystyle \frac{\mathrm{True} \mathrm{Positives} \left(\mathrm{TP}\right)+\mathrm{True} \mathrm{Negative} \left(\mathrm{TN}\right)}{\mathrm{Total} \mathrm{Number} \mathrm{of} \mathrm{Samples}}}$$

$$F1 Score=2\times {\displaystyle \frac{\mathrm{Precision}\times \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}}$$

$$ROC-AUC={{\displaystyle \int }}_0^1\mathrm{TPR} \left(\mathrm{True} \mathrm{Positive} \mathrm{Rate}\right) d\left(\mathrm{FPR} \left(\mathrm{False} \mathrm{Positive} \mathrm{Rate}\right)\right)$$

For the VDD classification, CatBoost produced the highest F1 score and was thus used to compute feature importance across 50 cross-validation runs. We averaged the importance ranks from each iteration and identified 20 top factors. From these, variables directly associated with vitamin D (age, residential area, dietary vitamin D intake) were excluded, leaving 17 key factors that were subsequently validated via stratified statistical analyses.

2.5. Statistical Analyses

The 17 key VDD-related factors, identified through machine learning, along with general characteristics, were analyzed based on vitamin D status (deficient or sufficient). Continuous variables are presented as means ± standard errors (SE), while categorical variables are presented as percentages. The reported proportions (%) were weighted to account for the complex survey design. To assess differences between vitamin D statuses, independent t-tests and Rao–Scott chi-square tests were applied to continuous and categorical variables, respectively.

The associations between blood 25(OH)D3 level and each of the continuous variables were analyzed using multivariate linear regression models. For the categorical variables (breakfast intake frequency, dietary supplement use, and urban residence), VDD risk was evaluated using multivariate logistic regression models, with the highest breakfast intake frequency, the non-use of dietary supplements, and urban residence as the reference groups, respectively. To minimize confounding effects and account for differences in covariate distribution between vitamin-D-deficient and sufficient groups [1,24,25,26], we employed three models in both linear and logistic regression analyses: (1) the unadjusted model provided crude beta coefficients (β) and SE values for continuous variables as well as crude odds ratios (ORs) and 95% confidence intervals (CIs) for categorical variables; (2) Model 1 was adjusted for age, BMI, and total energy intake; and (3) Model 2 included additional adjustments for household income, lifestyle factors (alcohol consumption, smoking status, and regular aerobic exercise), and dietary vitamin D intake.

All statistical analyses were performed using SPSS software (version 26; IBM, Armonk, NY, USA), with adjustments made for complex survey design effects [27]. A two-sided p-value less than 0.05 was considered statistically significant, and exact p-values were reported to facilitate interpretation.

3. Results

To classify vitamin D sufficiency and deficiency, we initially trained and evaluated six tree ensemble-based algorithms using various accuracy metrics. Among these, CatBoost demonstrated the highest performance in terms of precision, accuracy, F1 score, and area under the receiver operating characteristic curve (Figure 2); therefore, it was selected as the primary model for identifying key VDD-associated variables.

Using the CatBoost model, we performed 50 rounds of cross-validation to compute feature importance and subsequently averaged the rankings obtained from each iteration. This process identified 20 variables closely related to VDD from a machine learning perspective. The consistency of their rankings across the 50 iterations, visualized using a heatmap (Figure 3), confirmed the robustness of these findings. The 20 variables were dietary supplement use, age, blood urea nitrogen, urban residence, waist circumference, serum HDL cholesterol, urinary sodium, blood creatinine, folate intake, average daily sitting time, vitamin intake, fasting plasma glucose, SGOT, water intake, body weight, breakfast intake frequency, hemoglobin, region of residence, urinary creatinine, and blood red cell count. From these 20 variables, those either directly related to vitamin D or used for stratification, namely age, place of residence, and dietary vitamin D intake, were excluded, leaving 17 core variables that significantly influence serum vitamin D levels.

Thereafter, the top 17 VDD-associated variables identified through CatBoost machine learning were further analyzed by categorizing participants according to sex and age to assess differences based on vitamin D status. The percentage of participants with VDD (<20 ng/mL of serum 25[OH]D3) was 51.2%, 45.2%, 33.7%, and 22.7% in younger men (19–64 years), younger women (19–64 years), older men (≥65 years), and older women (≥65 years), respectively. Regarding general characteristics, the average age significantly varied with vitamin D status in younger men, younger women, and older men (Supplementary Tables S1 and S2). Among younger adults aged 19–64 years, the vitamin-D-sufficient group was significantly older than the vitamin-D-deficient group, while in older men, the vitamin-D-deficient group was older. Household income, BMI, and total energy intake did not significantly differ with vitamin D status across all age groups and sexes. Only in younger women were the current alcohol consumption and smoking rates significantly higher in the vitamin-D-deficient group than in the vitamin-D-sufficient group. In the vitamin-D-sufficient group, compared to those in the deficient group, women aged ≥ 65 years exhibited a higher rate of regular aerobic exercise, while men aged ≥ 65 years demonstrated higher daily vitamin D intake.

The top 17 factors for VDD, identified via CatBoost machine learning, were analyzed by comparing the vitamin D deficient and sufficient groups, categorized by sex and age (above and below 65 years) (

Table 1. Characteristics of the 17 variables identified through machine learning in participants aged 19–64 years.

	Men (n = 1424)			Women (n = 1895)
	VD Deficient (n = 729)	VD Sufficient (n = 695)	p Value *	VD Deficient (n = 857)	VD Sufficient (n = 1038)	p Value *
Breakfast intake frequency, %			<0.001			<0.001
5–7 times/wk	261 (33)	393 (52)		310 (33)	583 (54)
3–4 times/wk	96 (14)	78 (13)		121 (14)	138 (13)
1–2 times/wk	109 (16)	75 (12)		172 (21)	119 (11)
<1 time/wk	263 (37)	149 (23)		254 (32)	198 (21)
Dietary supplement use, %	390 (53)	528 (78)	<0.001	550 (63)	916 (89)	<0.001
Urban residence, %	632 (90)	528 (81)	<0.001	750 (91)	835 (85)	0.007
Blood urea nitrogen	13.4 ± 0.1	14.6 ± 0.1	<0.001	12.2 ± 0.1	13.4 ± 0.2	<0.001
Waist circumference, cm	89.1 ± 0.4	88.3 ± 0.4	0.13	77.8 ± 0.5	77.8 ± 0.4	0.96
Serum HDL cholesterol	50.6 ± 0.5	52.2 ± 0.6	0.03	63.7 ± 0.7	65.4 ± 0.5	0.04
Urinary sodium	115 ± 2	113 ± 2	0.45	104 ± 2	99.2 ± 1.8	0.06
Blood creatinine	0.91 ± 0.01	0.93 ± 0.01	0.005	0.67 ± 0.0	0.07 ± 0.01	0.006
Folate intake	314 ± 5	338 ± 7	0.006	264 ± 6	286 ± 6	0.005
Average daily sitting time, h	14.4 ± 0.8	15.9 ± 1.2	0.07	14.0 ± 1.0	14.0 ± 0.7	0.33
Fasting plasma glucose	102 ± 1	102 ± 1	0.74	95.4 ± 0.7	95.7 ± 0.8	0.76
SGOT	24.4 ± 0.6	24.8 ± 0.6	0.59	18.8 ± 0.4	20.3 ± 0.3	0.001
Water intake	1103 ± 26	1212 ± 26	0.002	950 ± 21	988 ± 19	0.14
Body weight	76.5 ± 0.5	75.3 ± 0.5	0.09	60.4 ± 0.5	58.5 ± 0.4	0.002
Hemoglobin	15.3 ± 0.0	15.2 ± 0.0	0.008	12.9 ± 0.0	13.1 ± 0.0	<0.001
Urinary creatinine	183 ± 4	160 ± 4	<0.001	142 ± 3	114 ± 3	<0.001
Red blood cell count	5.1 ± 0.0	5.0 ± 0.0	<0.001	4.4 ± 0.0	4.4 ± 0.0	0.07

Data are expressed as the mean ± SE for continuous variables or numbers (%) for categorical variables. The reported proportions (%) were weighted to account for the complex survey design. * Differences were determined via ANOVA for continuous variables or Rao–Scott chi-square tests for categorical variables. ANOVA, analysis of variance; HDL, high-density lipoprotein; SE, standard error; SGOT, serum glutamic–oxaloacetic transaminase; VD, vitamin D.

and

Table 2. Characteristics of the 17 variables identified through machine learning in participants aged ≥ 65 years.

	Men (n = 671)			Women (n = 831)
	VD Deficient (n = 226)	VD Sufficient (n = 445)	p Value *	VD Deficient (n = 222)	VD Sufficient (n = 609)	p Value *
Breakfast intake frequency, %			0.68			0.75
5–7 times/wk	211 (94)	419 (93)		197 (90)	545 (89)
3–4 times/wk	3 (1)	7 (3)		7 (3)	21 (4)
1–2 times/wk	5 (2)	4 (1)		6 (2)	18 (3)
<1 time/wk	7 (3)	15 (4)		12 (5)	25 (4)
Dietary supplement use, %	92 (44)	304 (70)	<0.001	124 (57)	467 (78)	<0.001
Urban residence, %	176 (84)	266 (70)	0.01	153 (78)	386 (74)	0.35
Blood urea nitrogen	16.9 ± 0.4	17.5 ± 0.3	0.15	16.7 ± 0.4	16.7 ± 0.3	1.00
Waist circumference, cm	89.5 ± 0.6	88.8 ± 0.5	0.34	86.9 ± 0.8	84.7 ± 0.5	0.02
Serum HDL cholesterol	50.5 ± 0.9	52.8 ± 0.8	0.052	57.1 ± 1.1	58.3 ± 0.7	0.35
Urinary sodium	125 ± 4	121 ± 3	0.48	119 ± 4	109 ± 2	0.03
Blood creatinine	1.01 ± 0.02	0.96 ± 0.01	0.06	0.74 ± 0.02	0.73 ± 0.01	0.28
Folate intake	379 ± 12	377 ± 9	0.91	318 ± 13	311 ± 10	0.63
Average daily sitting time, h	19.0 ± 1.9	16.2 ± 1.6	0.76	27.0 ± 3.0	18.8 ± 1.4	0.02
Fasting plasma glucose	113 ± 3	107 ± 1	0.02	108 ± 2	104 ± 1	0.07
SGOT	23.6 ± 0.8	24.5 ± 0.5	0.32	23.2 ± 0.5	24.2 ± 0.5	0.18
Water intake	856 ± 31	964 ± 33	0.02	744 ± 35	769 ± 25	0.52
Body weight	66.1 ± 0.7	66.7 ± 0.5	0.51	58.1 ± 0.8	56.9 ± 0.4	0.21
Hemoglobin	14.2 ± 0.1	14.2 ± 0.1	0.88	12.8 ± 0.1	12.9 ± 0.05	0.29
Urinary creatinine	104 ± 4	109 ± 3	0.50	79.7 ± 3.6	72.5 ± 1.9	0.08
Red blood cell count	1.59 ± 0.03	4.58 ± 0.02	0.65	4.23 ± 0.03	4.24 ± 0.02	0.80

Data are expressed as the mean ± SE for continuous variables or numbers (%) for categorical variables. The reported proportions (%) were weighted to account for the complex survey design. * Differences were determined via ANOVA for continuous variables or Rao–Scott chi-square tests for categorical variables. ANOVA, analysis of variance; HDL, high-density lipoprotein; SE, standard error; SGOT, serum glutamic–oxaloacetic transaminase; VD, vitamin D.

). In all age and sex groups, dietary supplement use (%) was higher in the vitamin-D-sufficient group (all p < 0.001). Additionally, urban residence (%) was significantly more prevalent in the vitamin-D-deficient group than in the sufficient group, except for women aged ≥ 65 years. Several variables, including breakfast intake frequency, blood urea nitrogen, serum HDL cholesterol, blood creatinine, folate intake, hemoglobin, and urinary creatinine, significantly varied with vitamin D status only in younger adults aged 19–64 years, but not in older adults aged ≥ 65 years. In both younger and older men, water intake was significantly higher in the vitamin-D-sufficient group than in the deficient group.

Subsequently, associations between blood 25(OH)D3 level and each of the 14 continuous variables among the top 17 associated factors were analyzed using multivariate linear regression models (

Table 3. Association between serum 25(OH)D3 level and 14 VDD-associated continuous variables identified through machine learning by age and sex.

Continuous Variables	19–64 Years	Model 1 *		Model 2 †		≥65 Years	Model 1 *		Model 2 †
		β	p-Value	β	p-Value		β	p-Value	β	p-Value
Blood urea nitrogen	Men	0.386	<0.001	0.398	<0.001	Men	0.297	<0.001	0.313	0.005
	Women	0.105	0.16	0.121	0.09	Women	0.179	0.03	0.160	0.06
Waist circumference	Men	−0.185	0.008	−0.190	0.01	Men	0.012	0.92	0.040	0.72
	Women	−0.048	0.56	−0.028	0.75	Women	−0.077	0.52	−0.116	0.39
Serum HDL cholesterol	Men	0.075	0.001	0.067	0.003	Men	0.041	0.25	0.044	0.25
	Women	0.042	0.02	0.041	0.03	Women	0.071	0.09	0.052	0.18
Urinary sodium	Men	−0.010	0.04	−0.009	0.08	Men	−0.016	0.12	−0.016	0.15
	Women	−0.017	0.004	−0.015	0.005	Women	−0.037	0.002	−0.037	0.004
Blood creatinine	Men	8.16	<0.001	8.16	<0.001	Men	−1.926	0.34	−1.97	0.31
	Women	4.93	0.04	5.19	0.02	Women	1.286	0.65	1.17	0.68
Folate intake	Men	0.002	0.32	0.001	0.62	Men	−0.001	0.82	−0.001	0.68
	Women	0.001	0.55	0.002	0.49	Women	−0.001	0.89	−0.001	0.79
Average daily sitting time	Men	0.010	0.37	0.002	0.98	Men	−0.016	0.42	−0.065	0.09
	Women	0.015	0.27	−0.061	0.34	Women	−0.026	0.24	−0.032	0.50
Fasting plasma glucose	Men	−0.023	0.003	−0.022	0.006	Men	−0.035	0.006	−0.033	0.009
	Women	−0.018	0.20	−0.012	0.39	Women	−0.008	0.75	−0.000	1.00
SGOT	Men	0.009	0.56	−0.004	0.79	Men	0.012	0.83	−0.008	0.89
	Women	0.039	0.14	0.032	0.18	Women	0.035	0.45	0.036	0.42
Water intake	Men	0.002	0.003	0.001	0.02	Men	0.003	0.006	0.003	0.02
	Women	0.001	0.02	0.002	0.02	Women	0.002	0.25	0.002	0.27
Weight	Men	0.086	0.06	0.084	0.08	Men	0.163	0.08	0.209	0.02
	Women	0.007	0.90	0.009	0.90	Women	0.047	0.70	−0.064	0.61
Hemoglobin	Men	−0.364	0.14	−0.280	0.26	Men	0.152	0.72	0.055	0.89
	Women	0.778	0.002	0.791	0.003	Women	0.157	0.73	0.392	0.43
Urinary creatinine	Men	−0.010	0.001	−0.009	0.002	Men	−0.009	0.26	−0.007	0.38
	Women	−0.014	<0.001	−0.014	0.001	Women	−0.029	0.009	−0.036	0.001
Red blood cells	Men	−0.973	0.18	−0.762	0.31	Men	−0.120	0.92	−0.228	0.84
	Women	−0.536	0.54	−0.412	0.65	Women	−0.414	0.75	0.398	0.77

This analysis was conducted using multiple regression analysis between serum 25(OH)D3 level and each variable identified through machine learning with adjustment. A table presenting the standard errors of β is prepared in Supplementary Table S3. * Model 1: adjusted for age, body mass index, and total energy intake. ^† Model 2: model 1 plus adjustments for household income, alcohol consumption, smoking, aerobic exercise, and vitamin D intake. 25(OH)D3, 25-hydroxyvitamin D3; HDL, high-density lipoprotein; SGOT, serum glutamic–oxaloacetic transaminase; VDD, vitamin D deficiency.

and Supplementary Table S3). Several variables significantly associated with serum 25(OH)D level differed by age group and sex in Model 2 after adjusting for age, BMI, total energy intake, household income, alcohol consumption, smoking, aerobic exercise, and vitamin D intake. In younger adults of both sexes, serum 25(OH)D level positively correlated with serum HDL cholesterol (men: β = 0.067, SE = 0.022, p = 0.003; women: β = 0.041, SE = 0.019, p = 0.03), blood creatinine (men: β = 8.16, SE = 1.95, p < 0.001; women: β = 5.19, SE = 2.27, p = 0.02), and water intake (men: β = 0.001, SE = 0.001, p = 0.02; women: β = 0.002, SE = 0.001, p = 0.02) but negatively correlated with urinary creatinine (men: β = −0.009, SE = 0.003, p = 0.002; women: β = −0.014, SE = 0.004, p = 0.001). In men across all age groups, blood urea nitrogen (19–64 years: β = 0.398, SE = 0.075; ≥65 years: β = 0.313, SE = 0.109) and water intake (19–64 years: β = 0.001, SE = 0.001; ≥65 years: β = 0.003, SE = 0.001) exhibited significantly positive associations with serum 25(OH)D level, while fasting plasma glucose displayed a negative association (19–64 years: β = −0.022, SE = 0.008, p = 0.006; ≥65 years: β = −0.033, SE = 0.012, p = 0.009). On the other hand, urinary sodium and creatinine levels yielded significantly negative correlations with serum 25(OH)D levels in both younger and older women.

Multivariate logistic regression models were used to ascertain whether each of the three categorical variables among the top 17 factors was associated with VDD according to sex and age (

Table 4. Risk of VDD according to each VDD-associated categorical variable (breakfast intake frequency, dietary supplement intake, or urban residence) identified through machine learning by age and sex.

Categorical Variables	Sex (Age)		Model 1 *		Model 2 †
			OR	95% CI	OR	95% CI
Breakfast intake frequency	Men (19–64)	<1 time/wk	2.07	1.53, 2.79	1.89	1.38, 2.59
		1–2 times/wk	1.76	1.16, 2.66	1.65	1.09, 2.51
		3–4 times/wk	1.30	0.88, 1.92	1.25	0.84, 1.85
		5–7 times/wk	1 (ref)		1 (ref)
	Women (19–64)	<1 time/wk	1.61	1.22, 2.11	1.47	1.11, 1.96
		1–2 times/wk	2.04	1.42, 2.95	2.02	1.36, 2.99
		3–4 times/wk	1.27	0.92, 1.76	1.23	0.88, 1.72
		5–7 times/wk	1 (ref)		1 (ref)
	Men (≥65)	<1 time/wk	0.82	0.30, 2.26	0.92	0.31, 2.69
		1–2 times/wk	2.31	0.52, 10.2	2.19	0.56, 8.52
		3–4 times/wk	0.74	0.17, 3.25	0.79	0.17, 3.64
		5–7 times/wk	1 (ref)		1 (ref)
	Women (≥65)	<1 time/wk	1.15	0.56, 2.38	1.09	0.48, 2.46
		1–2 times/wk	0.73	0.25, 2.15	0.76	0.24, 2.37
		3–4 times/wk	0.63	0.21, 1.89	0.82	0.27, 2.50
		5–7 times/wk	1 (ref)		1 (ref)
Dietary supplement use	Men (19–64)	No	3.06	2.21, 4.23	3.13	2.23, 4.41
		Yes	1 (ref)		1 (ref)
	Women (19–64)	No	3.90	2.94, 5.16	3.71	2.75, 4.99
		Yes	1 (ref)		1 (ref)
	Men (≥65)	No	2.91	1.91, 4.43	3.26	2.11, 5.03
		Yes	1 (ref)		1 (ref)
	Women (≥65)	No	2.77	1.90, 4.05	2.31	1.51, 3.54
		Yes	1 (ref)		1 (ref)
Urban residence	Men (19–64)	Urban	1.80	1.11, 2.90	1.94	1.19, 3.16
		Rural	1 (ref)		1 (ref)
	Women (19–64)	Urban	1.67	1.06, 2.62	2.00	1.26, 3.18
		Rural	1 (ref)		1 (ref)
	Men (≥65)	Urban	2.33	1.36, 4.02	2.88	1.67, 4.97
		Rural	1 (ref)		1 (ref)
	Women (≥65)	Urban	1.26	0.92, 1.74	1.26	0.86, 1.84
		Rural	1 (ref)		1 (ref)

This analysis was conducted using multivariate logistic regression models to assess the risk of VDD with adjustments. * Model 1: adjusted for age, body mass index, and total energy intake. ^† Model 2: model 1 plus adjustments for household income, alcohol consumption, smoking, aerobic exercise, and vitamin D intake. CI, confidence interval; OR, odds ratio; VDD, vitamin D deficiency.

). In Model 2 (adjusted for confounding factors), the non-use of dietary supplements had significantly higher odds of VDD than supplement use in younger men (OR, 3.13; 95% CI, 2.23–4.41), younger women (OR, 3.71; 95% CI, 2.75–4.99), older men (OR, 3.26; 95% CI, 2.11–5.03), and older women (OR, 2.31; 95% CI, 1.51–3.54). In addition, urban residence (%) exhibited significantly higher odds of VDD than rural residence in younger men (OR, 1.94; 95% CI, 1.19–3.16), younger women (OR, 2.00; 95% CI, 1.26–3.18), and older men (OR, 2.88; 95% CI, 1.67–4.97). In younger adults of both sexes, consuming breakfast less than once a week or 1–2 times a week was associated with significantly higher odds of VDD than consuming breakfast 5–7 times a week. However, older adults did not display a significant association between breakfast intake frequency and VDD. A summarized heatmap of the results outlined in Table 3 and Table 4 is shown in Figure 4.

4. Discussion

Our findings reveal that VDD remains a significant public health concern in South Korea, particularly among younger adults (45.2–51.2%). Using the most recent KNHANES IX-1 data, this study identified several novel VDD-associated factors in the Korean population. Machine learning techniques initially screened the top 17 factors, which were then further analyzed by age and sex using statistical methods. Among these, dietary supplement use showed a consistent and strong association with VDD across all age and sex groups. In younger adults of both sexes, HDL cholesterol, blood and urinary creatinine, water intake, urban residence, and breakfast intake frequency were significantly linked to VDD. Additionally, several factors, including blood urea nitrogen and fasting plasma glucose in men and urinary sodium in women, were associated with serum 25(OH)D levels in a sex-dependent manner.

Machine learning has increasingly been applied in health data analytics, demonstrating its capability to uncover complex relationships and improve diagnostic accuracy across diverse health domains. Several studies have applied machine learning techniques to cross-sectional models using KNHANES data to identify factors associated with specific diseases [28,29,30,31]. For instance, metabolic syndrome studies utilizing deep learning and ensemble methods have identified key contributors such as dietary intake, triglycerides, and waist-to-height ratio [28,29]. Similarly, machine learning analyses of cardiovascular disease risk factors, including LightGBM and XGBoost, have highlighted age, hypertension, and BMI as major determinants [30]. Moreover, machine learning has been applied to conditions such as osteoporosis and obesity, demonstrating its capacity to integrate biochemical, lifestyle, and genetic data, uncover complex associations, and provide stratified insights into health conditions [28,31]. In the context of VDD, most studies have primarily relied on traditional statistical methods [32,33,34], with limited application of machine learning. A recent study on VDD using NHANES data applied machine learning techniques such as Gradient Boosting Machines, Neural Networks, and XGBoost; however, it focused on classifying and predicting VDD rather than finding VDD-related factors [35]. Therefore, this study may be the first to integrate machine learning and statistical methods to comprehensively identify key VDD-related factors, offering a deeper understanding of its determinants.

In this study, the non-use of dietary supplements emerged as the most significant factor associated with VDD across all sex and age subgroups. Regarding dietary supplements use, the KNHANES IX-1 assessed experiences of taking supplements for >2 weeks within the preceding year, without detailed investigation into specific supplement use. Therefore, accurately assessing vitamin D intake from supplement use proves challenging; however, among Koreans, the associations with VDD might be attributable to considerable vitamin D intake via dietary supplements. Several previous studies have established vitamin D as a top dietary supplement consumed by Koreans [36,37]. According to an analysis based on 2015 KNHANES data, multivitamin mineral supplements constituted the most frequently consumed dietary supplement in both men and women, with a rate of 89.6 per 1000 people, while vitamin D ranked seventh [37]. As numerous multivitamin supplements contain vitamin D, cumulative intake from multiple dietary supplements has been reported as a potential reason for exceeding the upper limit of vitamin D [38]. Additionally, a study on Koreans aged >10 years found that the elderly had the highest serum 25(OH)D levels, despite their low dietary vitamin D intake, suggesting supplementation as a contributing factor [34,39]. In particular, vitamin D was the most frequently supplemented nutrient during the first and second waves of the COVID-19 pandemic owing to its beneficial role in supporting immune defense against viral infections [10,40]. Therefore, the close association between VDD and dietary supplement use in this study is potentially attributable to high vitamin-D-related supplement intake in the Korean population.

Interestingly, certain blood factors were significantly associated with vitamin D only in younger adults, including its positive correlation with serum HDL cholesterol, aligning with previous findings [41,42,43]. A study analyzing 28,084 NHANES adults aged 20–59 years or yielded a significantly positive correlation between HDL cholesterol and vitamin D (25[OH]D) level [41]. This positive association between HDL cholesterol and vitamin D levels may be related to the protective effects of HDL cholesterol on cardiovascular health, suggesting that vitamin D potentially supports cardiovascular health through its role in lipid metabolism. This highlights the potential cardiovascular benefits of maintaining optimal vitamin D levels [43].

The current study also found younger adults residing in rural areas to have higher vitamin D levels than their urban counterparts, possibly because of greater outdoor activity in rural settings, resulting in more sunlight exposure and, therefore, higher vitamin D synthesis. Wakayo et al. [44] reported a higher prevalence of vitamin D deficiency in urban (61.8%) than rural (21.2%) Ethiopian schoolchildren. However, a study in western Ireland found lower serum 25(OH)D levels in rural adults year-round [45], likely due to geographic differences. In addition, we found an association between breakfast intake and vitamin D in younger adults, consistent with previous studies. Research in Brazilian adolescents and US Army recruits highlighted that regular breakfast consumption improves vitamin D levels and reduces VDD risk [46,47].

Creatinine, a byproduct of creatine phosphate breakdown in muscles, is a marker of skeletal muscle mass under stable kidney function [48,49]. When kidney function declines, blood creatinine levels increase, while urinary creatinine decreases due to reduced filtration capacity [50]. In younger adults, blood creatinine positively correlated with serum 25(OH)D, whereas urinary creatinine showed a negative correlation. Vitamin D has been linked to kidney function, with studies reporting its negative association with albuminuria and potential benefits in CKD management [51,52]. Moreover, vitamin D may influence muscle creatinine production, as clinical trials have shown increased serum creatinine levels following vitamin D receptor activation, independent of kidney function [53,54].

Several VDD-associated variables, such as blood urea nitrogen, fasting plasma glucose, and urinary sodium exhibited sex-specific differences. As blood urea nitrogen and urinary sodium reflect kidney health, these findings suggest a potential link between vitamin D status and kidney function. Furthermore, fasting glucose and urinary sodium are related to type 2 diabetes and hypertension, conditions in which vitamin D reportedly influences insulin secretion, sensitivity, and β-cell function while exerting anti-inflammatory effects [7]. Further research is needed to explore these sex-specific associations and potential therapeutic implications. Lastly, dietary folate intake showed no significant association with VDD when stratified by sex and age, despite being identified as a key factor by the CatBoost model. This discrepancy may result from interaction effects or reduced statistical power due to stratification. The “vitamin D–folate hypothesis” suggests an evolutionary link between skin pigmentation, vitamin D synthesis, and folate preservation, as UVR both stimulates vitamin D production and degrades folate. Additionally, in vitro studies indicate that vitamin D upregulates folate transporters, further supporting this interplay [55,56,57]. While these findings imply a potential connection between vitamin D and folate, further research is needed to clarify their relationship.

This study has several limitations. First, its cross-sectional design prevents causal inferences, and reverse causation cannot be ruled out. Second, key determinants of 25(OH)D levels, such as sun exposure time and seasonal effects [58,59,60,61,62], were not adjusted for due to data limitations in KNHANES IX-1. While we indirectly assessed sun exposure by comparing urban and rural residents, this approach does not fully capture its effects. Third, we used a standard cutoff to classify participants as vitamin D deficiency as <20 ng/mL [19], defining groups by a relatively narrow margin may skew causative interpretations. Fourth, self-reported lifestyle factors may introduce recall bias despite the availability of objective laboratory data. Fifth, key variables like medication use and surgical history lacked detail, limiting further analysis. Sixth, the dataset does not specify supplement type or dosage, preventing precise evaluation of their impact on vitamin D levels. Lastly, the study may be subject to statistical power limitations (Type II error), particularly for older age groups with a smaller sample size, potentially affecting the detection of significant associations in stratified analyses. Despite these limitations, this study provides a comprehensive analysis of VDD-related factors in a large national sample, leveraging machine learning for robust feature selection. Future research should incorporate sun exposure, seasonal variations, alternative criteria for differentiating between vitamin D deficiency and sufficiency, supplement details, and detailed dietary assessments for a more precise understanding of VDD.

5. Conclusions

VDD remains a prevalent health issue in South Korea. Through an integrative approach combining machine learning and statistical analyses, several key factors associated with VDD were identified, including dietary supplement use, urinary creatinine levels, urban residence, and breakfast frequency. In this study, machine learning models capable of handling complex correlations among variables were employed to screen the most relevant factors, facilitating the analysis of a comprehensive set of 583 variables. As bioanalytical and data processing technologies continue to advance, future research will need to consider an even greater number of factors, further emphasizing the importance of machine learning approach. Furthermore, the inclusion of new data sources (i.e., KNHANES IX-1) and the implementation of age- and sex-stratified statistical analyses are also key strengths of this study. This research highlights the significance of integrating machine learning with statistical analysis, demonstrating its potential for more comprehensive and insightful investigations.