Predicting Type 2 Diabetes Using Baseline and Longitudinal Changes in Lifestyle and Clinical Markers: A Machine Learning Approach

Predicting Type 2 Diabetes Using Baseline and Longitudinal Changes in Lifestyle and Clinical Markers: A Machine Learning Approach | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Predicting Type 2 Diabetes Using Baseline and Longitudinal Changes in Lifestyle and Clinical Markers: A Machine Learning Approach Kim Taegyun, Jiwon Do, Hye Won Woo This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7902454/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 8 You are reading this latest preprint version Abstract Most type 2 diabetes (T2D) prediction models rely on static baseline measurements and often include diagnostic glycemic markers, limiting their ability to capture temporal risk evolution and creating circular reasoning. This study developed a machine learning framework that systematically integrates baseline measurements with longitudinal interval changes to predict incident T2D. Using the Ansan-Ansung cohort of the Korean Genome and Epidemiology Study (KoGES; 2001–2018), we included 7,510 initially diabetes-free participants in this prospective analysis. The framework jointly modeled static variables and 2-year interval changes in lifestyle, anthropometric, and biochemical markers using XGBoost, Random Forest, LGBM, logistic regression, neural networks, and ensemble methods. Principal component analysis addressed multicollinearity. Diagnostic glycemic markers (fasting glucose, HbA1c) were excluded to ensure genuine risk prediction. The ensemble model achieved AUROC 0.763, with XGBoost (0.752) and LGBM (0.750) showing comparable performance. SHAP analysis identified changes in C-reactive protein (▲CRP) and body mass index (▲BMI), together with baseline triglycerides, as the most influential predictors. Examination of decision tree structures revealed clinically meaningful and biologically plausible thresholds (e.g., BMI < 25.6 kg/m²). The resulting ensemble model was implemented through the Multi-Domain Simulation Interface (MDSi) framework, enabling population-level inference across lifestyle, anthropometric, and metabolic domains. Overall, change variables contributed more strongly than static measures, suggesting that accelerated physiological shifts precede the onset of T2D. By capturing dynamic metabolic trajectories rather than static risk profiles, this framework differentiates true risk prediction from early disease detection, enabling clinically interpretable prediction with substantial potential for preventive interventions before diagnostic thresholds are reached. Health sciences/Biomarkers Biological sciences/Computational biology and bioinformatics Health sciences/Diseases Health sciences/Endocrinology Health sciences/Health care Health sciences/Medical research Health sciences/Risk factors Figures Figure 1 Figure 2 1. Introduction Type 2 Diabetes (T2D) remains a global health burden, affecting 451 million people worldwide in 2017 and projected to reach 693 million by 2045 1 . The development of T2D is typically characterized by either a gradual accumulation of risk factors from middle adulthood, a sudden deterioration in health shortly before diagnosis, or a combination of both trajectories 2 , 3 . In recent years, there has been growing interest in leveraging machine learning techniques to improve the prediction of T2D onset 4 – 6 . Current machine learning approaches for T2D prediction face two critical limitations. First, these models predominantly rely on cross-sectional or static baseline data, failing to capture the temporal dynamics of risk factor evolution 4 , 5 , 7 . This is problematic because diabetes risk accumulates over time—an individual with elevated risk factors for 10 years has substantially different risk than someone with identical readings for one year, yet most models cannot distinguish between these scenarios 7 . Second, many existing models incorporate diagnostic glycemic markers (glucose, HbA1c) as input features 8 , 9 , creating a logical flaw of using disease-defining variables to predict the disease itself. These limitations collectively undermine model validity: current approaches neither capture temporal risk accumulation nor distinguish between genuine prediction and early disease detection. While machine learning shows promise in healthcare, there remains a critical gap in understanding how temporal changes in lifestyle factors and biological markers contribute to diabetes risk progression. To address these limitations, this study applies multiple machine learning approaches to develop prospective prediction models for T2D onset using both static measurements and interval changes between consecutive time points in lifestyle factors and biological markers. Using data from the Ansan and Ansung cohort of the Korean Genome and Epidemiology Study (KoGES), we exclude changes in blood glucose and HbA1c to avoid circular reasoning. To ensure transparency in clinical decision support, we apply explainable AI methods to interpret how individual features contribute to risk predictions. 2. Methods 2.1. Study population Data were obtained from the Ansan–Ansung cohort of the Korean Genome and Epidemiology Study (KoGES), a 10-wave longitudinal cohort study of Korean adults aged 40 years or older at baseline (2001–2002), with biannual follow-ups conducted through 2019–2020. The Ansan–Ansung study is one of the KoGES population-based cohort studies involving community-dwelling individuals who participated in health examinations. It comprises men and women aged 40–69 years at baseline, residing in Ansan (an urban area) and Ansung (a rural area). All participants voluntarily participated and provided written informed consent. The study was conducted in accordance with the Declaration of Helsinki and was approved by the ethics committee of the Korean Health and Genome Study, supported by the Korea National Institute of Health (KNIH). Further details on the KoGES design and sampling methods have been described in previous reports 10 . For the purpose of this prospective study, among the 10,038 participants in the baseline survey, we excluded 1,352 (13.5%) participants who met the American Diabetes Association (ADA) criteria for T2D 11 : use of anti-diabetic medications or insulin, fasting blood glucose (FBG) ≥ 126 mg/dL (7.0 mmol/L), 2-hour blood glucose ≥ 200 mg/dL (11.1 mmol/L) during an oral glucose tolerance test (OGTT), or glycated hemoglobin (HbA1c) ≥ 6.5%. Furthermore, we excluded those with a history of cardiovascular disease, cerebrovascular disease, or cancer at baseline (n = 439) to eliminate potential confounding effects from these more severe conditions that could alter lifestyle behaviors and medication use. We also excluded participants who did not attend follow-up visits (n = 737). As a result, 7,510 participants (3,565 men and 3,945 women) were included in the final analysis. 2.2. Ascertainment of diabetes incidence Participants underwent nine biennial follow-up examinations from 2003–2004 to 2019–2020. At each visit, FBG, HbA1c, and 2-hour blood glucose levels after a 75g OGTT were measured. Using the same ADA criteria applied at baseline, incident T2D was assessed at each follow-up visit 11 . The detailed criteria for T2D identification are described in section 2.4.1 . 2.3. Assessment of lifestyle factors and biomarkers All predictor variables were collected at each follow-up visit by trained interviewers following a standardized protocol. Interviewers were trained to administer the structured questionnaire based on a standardized manual 10 . Predictor variables were categorized into three domains: lifestyle and socioeconomic factors, anthropometric and physical measurements, and metabolic and biochemical markers. Lifestyle and socioeconomic factors included smoking status (current, former, or never), pack-years (for current smokers), alcohol drinking status and alcohol consumption (categorized as current, former, or never; total alcohol intake was assessed only among current drinkers), physical activity (defined as moderate-intensity activity ≥ 3 times/week, ≥ 30 minutes/session), water intake frequency, urinary frequency, residential area (Ansan or Ansung), education level (high school graduate: yes or no), parental history of diabetes (yes or no), and housing type (no housing, owner-occupied, deposit-based or monthly rental, and other). Economic indicators included monthly average income, medical expenses, and expenditures on health foods. Anthropometric and physical measurements were obtained by trained health professionals using standardized procedures. Body weight was measured to the nearest 0.1 kg with participants wearing light indoor clothing without shoes. Height was measured to the nearest 0.1 cm. Body mass index (BMI) was calculated as weight (kg) divided by height squared (m²). Waist circumference (WC) was measured at the narrowest part between the lower rib and the iliac crest to the nearest 0.1 cm. Blood pressure was measured three times in the right upper arm in the supine position using a standard mercury sphygmomanometer, and the average of the second and third readings was used for analysis. Metabolic and biochemical markers were assessed from blood samples collected after overnight fasting. All samples were analyzed at a central laboratory (Seoul Clinical Laboratories, Seoul, Korea). Serum glucose, triglycerides, total cholesterol, HDL-cholesterol, liver enzymes (AST, ALT), kidney function markers (blood urea nitrogen [BUN], serum creatinine), albumin, and inflammatory marker (high-sensitivity C-reactive protein [hsCRP]) were measured using an automated chemistry analyzer (ADVIA 1650; Siemens, NY, USA). Complete blood count was used to determine white blood cell count (WBC), red blood cell count (RBC), platelet count (PLT), hematocrit (Hct), and hemoglobin (Hb). For measurements obtained prior to the introduction of ADVIA 1650 (September 2002), validated conversion equations were applied to ensure compatibility across instruments and consistency over time. Among all predictor variables, sex, residential area, education level, parental history of diabetes, and housing type were considered fixed baseline predictors. Missing values in the dataset were handled through imputation, where continuous variables were imputed using mean values and categorical variables were imputed using mode values to maintain data completeness for subsequent machine learning analysis. 2.4. Feature engineering and Model Development 2.4.1 Target Variable Definition To enhance predictive precision and account for potential confounding effects of medication on clinical parameters, we defined the target outcome using a laboratory parameter-based definition exclusively. We established T2D onset identification solely through objective measurements (FBG ≥ 126 mg/dL, 2-hour plasma glucose ≥ 200 mg/dL after OGTT, or HbA1c ≥ 6.5%) without including medication-based diagnoses. This approach enabled analysis of natural disease biomarker progression without pharmacological confounding, as anti-diabetic medications can normalize glucose levels, alter other clinical parameters, and prompt behavioral changes. For participants who initiated anti-diabetic medications or insulin therapy during the follow-up period, we censored their data at the last visit where they remained medication-free and maintained non-diabetic status. This censoring approach ensured that our analysis captured only the natural disease progression trajectory, avoiding potential bias introduced by treatment effects on parameters. 2.4.2 Change-Based Feature Engineering To capture temporal patterns in T2D onset among initially non-diabetic individuals, we utilized both static measurements (observed values at each visit) and generated change-based features representing interval changes in variables between consecutive visits. For participants who developed T2D, if T2D was diagnosed at visit Vₙ, the interval change (Vₙ-Vₙ₋₁) was considered indicative of the transition to diabetes, while all previous interval changes (V₂-V₁, V₃-V₂, ..., Vₙ₋₁-Vₙ₋₂) from the same participant represented normal variation. For example, if a participant was diagnosed at V₄, the change V₄-V₃ would represent diabetes-related transition, while changes V₂-V₁ and V₃-V₂ would represent normal fluctuations. This approach yielded 1,796 data points corresponding to visits immediately preceding diabetes diagnosis and 45,849 data points corresponding to pre-diagnostic visits from participants who remained non-diabetic throughout follow-up. The model was trained to predict the probability of T2D onset at visit Vₜ based on observed features at visit Vₜ₋₁ and their changes between Vₜ₋₁ and Vₜ. This approach enabled the model to identify patterns from both current visit measurements and their interval changes that distinguish imminent diabetes onset from normal fluctuation. 2.4.3 Dimensionality Reduction Using Principal Component Analysis (PCA) Most ML-based studies overlook the effect of multicollinearity on risk factor analysis 12 . Although multicollinearity does not affect the predictive power of ML models, variables with high collinearity offset the importance of each other, consequently leading to erroneous evaluation of variable importance 13 , 14 . Correlation analysis of our dataset revealed varying degrees of multicollinearity across different variable domains ( Supplementary Fig. 1 ), with metabolic and biochemical markers showing the strongest intercorrelations, followed by anthropometric measurements and lifestyle factors. To address multicollinearity issues, we applied Principal Component Analysis (PCA) transformation, which converts correlated input variables into orthogonal components. The input features were categorized into three biologically relevant domains: (1) lifestyle and socioeconomic factors, (2) anthropometric and physical measurements, and (3) metabolic and biochemical markers. PCA was applied separately within each domain to generate interpretable principal components while preserving domain-specific biological relationships. We employed three distinct analytical strategies to evaluate the optimal feature representation: (1) Original Features Analysis - using only the original variables without dimensionality reduction; (2) Hybrid Analysis - combining original variables with PCA-derived components; and (3) PC-Based Analysis - using exclusively the extracted principal components as predictors. This comparative approach enabled us to assess the explanatory power of transformed components versus their additive value when combined with original variables. Additionally, we evaluated model performance using domain-specific feature sets (lifestyle, anthropometric, biochemical) as well as combined non-blood-based factors (lifestyle + anthropometric) to identify the most informative feature combinations. 2.4.4 Model Development 2.4.4.1 Baseline Model We employed logistic regression (LR) as the baseline model to establish a performance benchmark and assess the linear predictability of T2D onset. Logistic regression was selected for its interpretability and established use in clinical risk prediction, following the standard formulation: $$\:Logit\left(p\right)=\:{\beta\:}_{0}+\:\sum\:_{i=1}^{n}{\beta\:}_{i\:}{x}_{i}$$ This baseline approach enabled objective evaluation of whether improvements observed in more complex models constituted meaningful enhancements, while providing interpretable coefficients with corresponding statistical significance for each predictor. 2.4.4.2 Advanced Machine Learning Models Figure 1 illustrates the comprehensive machine learning pipeline developed for T2D prediction. The framework consists of three main phases: learning, validation, and inference. In the learning phase, pre-processed features from the ASAS cohort undergo model training using six different algorithms, which are then combined into an ensemble model. We implemented five machine learning algorithms to capture complex, non-linear relationships in T2D prediction: Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Artificial Neural Network (ANN), and Recurrent Neural Network (RNN). Detailed model architectures, hyperparameter settings, and training procedures are provided in Supplementary Methods . Tree-based models (RF, XGBoost, LightGBM) were selected for their ability to capture non-linear relationships and feature interactions through recursive data partitioning 15 – 17 . These models construct decision trees by selecting optimal splits based on impurity reduction criteria, effectively handling categorical variables and missing values without requiring feature scaling 18 . Neural network models (ANN, RNN) were implemented to learn complex non-linear patterns through weighted transformations across multiple layers. ANN captures high-dimensional feature interactions through hidden layers 19 , while RNN was specifically designed to process sequential data by maintaining temporal dependencies through recurrent connections, enabling the model to capture patterns across time steps 20 . This framework illustrates the complete pipeline from model training using Ansan–Ansung cohort data (Learning) to inference and generalization on population-level data (Inference). Six machine learning algorithms (logistic regression, random forest, XGBoost, LightGBM, RNN, and ANN) were integrated into an ensemble model and applied through the Multi-Domain Simulation Interface (MDSi) framework to assess diabetes risk across lifestyle, anthropometric, and metabolic domains. 2.4.4.3 Ensemble Model Implementation We implemented a soft voting ensemble that combines predictions from all individual models by averaging their predicted probabilities. This approach integrates diverse model strengths while mitigating individual model limitations through complementary effects, particularly beneficial given the different architectural approaches and feature processing mechanisms of the base models. 2.5. Model Evaluation, Interpretability, and Feature Importance The dataset exhibited severe class imbalance, with positive cases (T2D onset) representing only 3.9% of total instances. To address this imbalance, we applied post-split over-sampling with replacement within each dataset partition—training (80%), validation (10%), and test (10%)—after data splitting. Class imbalance is common in epidemiological data, making accurate performance assessment challenging without balancing technique 21 . Over-sampling of the minority class was performed independently within each subset to achieve a balanced 1:1 class ratio while preventing information leakage and overfitting risk associated with pre-split oversampling 22 . Model performance was evaluated using accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUROC) to comprehensively assess both overall and minority class performance. To assess model interpretability and feature importance, we employed model-specific approaches: regression coefficients and p-values for Logistic Regression, feature importance values for tree-based models (RF, XGBoost, LightGBM), and predictive performance evaluation for neural network models (ANN, RNN). The ensemble approach combined predictions from all individual models through soft voting to enhance overall predictive accuracy. Detailed procedures are provided in the Supplementary Methods . 3. Results 3.1. Characteristics of the study participants As shown in Table 1 , a total of 7,510 participants were included in the final analysis, with 5,505 (73.3%) remaining non-diabetic and 2,005 (26.7%) developing incident T2D during follow-up. Among those who developed T2D, 1,796 (89.6%) were screen-detected cases identified through follow-up laboratory testing and 209 (10.4%) were medication-detected cases (for whom only pre-diabetic health data were included in the model). The study population consisted of middle-aged Korean adults (mean age 51.5 ± 8.7 years) with balanced sex distribution and equal urban-rural representation. Lifestyle and socioeconomic factors revealed that future T2D cases had higher medical expenses, with increased prevalence of family diabetes history (9.4% vs. 7.1%) and current smoking (28.1% vs. 24.6%). Anthropometric and physical measurements demonstrated significantly elevated baseline values in the T2D group, including higher BMI (25.1 ± 3.1 vs. 24.1 ± 3.0 kg/m²), waist circumference (84.9 ± 8.5 vs. 81.1 ± 8.6 cm), and blood pressure (123.7/80.9 vs. 118.9/79.0 mmHg). Metabolic and biochemical markers showed a distinct pre-diabetic profile in future T2D cases, with elevated glucose parameters even within normal ranges (FBG: 86.3 vs. 81.4 mg/dL; HbA1c: 5.7 vs. 5.48%), adverse lipid profiles (triglycerides: 180.2 vs. 145.2 mg/dL; HDL-cholesterol: 43.1 vs. 45.6 mg/dL), and higher inflammatory markers (hsCRP: 0.24 vs. 0.21 mg/dL), indicating subtle baseline metabolic differences that preceded future diabetes development, despite all participants being non-diabetic at study entry. Table 1. Baseline characteristics of the study Baseline variables * Total Consistently non-diabetic individuals T2D cases detected through follow-up testing T2D cases identified through medication initiation Participants, N 7,510 5,505 1,796 209 Lifestyle and socioeconomic factors Residential area (Ansung, %) 51.0 52.9 46.6 56.0 Male, % 47.5 46.4 41.6 51.5 Age, years 51.5 ± 8.7 51.2 ± 8.0 52.1 ± 5.5 54.2 ± 8.7 High school graduate 1 , % 13.7 13.7 14.4 8.7 Type of home ownership (Own, %) 83.0 82.7 83.8 82.3 Average monthly income, ten thousand won 1,825 ± 1,401 1,838 ± 1,401 1,822 ± 1,393 1,519 ± 1,411 Medical expenses, ten thousand won 42.0 ± 57.6 41.1 ± 56.5 43.9 ± 60.0 49.3 ± 65.1 Expenditures on health foods, ten thousand won 31.1 ± 56.8 31.4 ± 57.2 29.9 ± 54.6 35.7 ± 64.0 Family history of diabetes in parents, % 7.6 7.1 9.4 8.1 Regular exercise 2 , % 59.6 60.2 58.7 53.9 Current smoker, % 25.5 24.6 28.1 26.8 Pack-years (for current smokers), y 6.0 ± 13.2 5.7 ± 12.9 6.9 ± 13.9 6.7 ± 14.5 Current drinker, % 48.4 47.7 51.5 40.1 Alcohol consumption, g/d 9.2 ± 21.4 8.9 ± 20.9 10.6 ± 22.8 7.5 ± 21.2 Recent increase in frequency of drinking water, % 13.5 13.8 12.5 15.9 Recent increase in urination frequency, % 19.0 18.7 19.3 25.4 Anthropometric and physical measurements Body weight, kg 62.7 ± 10.1 62.0 ± 9.9 64.7 ± 10.3 65.5 ± 10.7 Waist circumference, cm 82.1 ± 8.7 81.3 ± 8.6 84.9 ± 8.5 84.2 ± 8.5 Body Mass Index, kg/m 2 24.5 ± 3.1 24.1 ± 3.0 25.1 ± 3.1 25.9 ± 3.2 Systolic blood pressure (SBP), mmHg. 120.3 ± 17.9 118.9 ± 17.5 123.7 ± 18.3 127.6 ± 18.4 Diastolic Blood Pressure (DBP), mmHg. 79.9 ± 11.4 79.0 ± 11.3 80.9 ± 11.2 84.4 ± 11.4 Metabolic and biochemical markers Fasting blood glucose (FBG), mg/dL 82.7 ± 8.5 81.4 ± 7.7 86.3 ± 9.7 87.0 ± 9.1 A haemoglobin A1C (HbA1C), % 5.54 ± 0.35 5.48 ± 0.32 5.7 ± 0.34 5.8 ± 0.34 Triglycerides (TG), mg/dL 155.0 ± 95.2 145.2 ± 87.0 180.2 ± 102.6 198.9 ± 160.4 Total cholesterol (Tchl), mg/dL 189.3 ± 34.1 187.7 ± 34.2 193.2 ± 33.5 197.1 ± 33.8 HDL-cholesterol, mg/dL 44.9 ± 10.1 45.6 ± 10.2 43.1 ± 9.4 42.7 ± 10.6 Aspartate transferase (AST), U/L 29.2 ± 16.9 28.6 ± 16.6 30.8 ± 18.2 29.0 ± 9.3 Alanine aminotransferase (ALT), U/L 27.1 ± 22.3 25.6 ± 19.9 31.3 ± 28.3 29.9 ± 16.6 Blood urea nitrogen (BUN), mg/dL 14.3 ± 3.6 14.2 ± 3.6 14.5 ± 3.7 14.6 ± 3.6 Serum creatinine, mg/dL 0.84 ± 0.19 0.83 ± 0.19 0.85 ± 0.19 0.82 ± 0.16 High-sensitivity C-reactive protein (hsCRP), mg/dL 0.22 ± 0.52 0.21 ± 0.52 0.24 ± 0.48 0.32 ± 0.83 White blood cell count (WBC), 10 3 /µL 6.48 ± 1.79 6.38 ± 1.79 6.7 ± 1.8 6.9 ± 1.9 Red blood cell count (RBC), 10 3 /µL 4.41 ± 0.46 4.39 ± 0.45 4.48 ± 0.42 4.42 ± 0.48 Platelet count (PLT), 10 3 /µL 265.9 ± 62.8 264.5 ± 62.1 269.6 ± 64.4 271.3 ± 66.0 Hematocrit (Hct), % 40.9 ± 4.5 40.8 ± 4.5 41.5 ± 4.6 41.1 ± 4.5 Hemoglobin (Hb), g/dL 13.6 ± 1.8 13.5 ± 1.6 13.8 ± 1.6 13.7 ± 1.5 * The values are expressed as mean ± SD for continuous variables or percentage for categorical variables. 1 High school graduate (≥ 12 years of education). 2 Regular exercise (≥ 3 times/week and ≥ 30 min/session). 3.2. Model Performance Comparison To demonstrate the circular nature of including diagnostic glycemic markers, supplementary analysis showed AUROC values exceeding 0.9 Supplementary Table 1 when blood glucose and HbA1c were included as predictors. Supplementary Figure 2 illustrates the SHAP analysis revealing that glucose-related variables (change in 2-hour OGTT glucose, 2-hour OGTT glucose level, HbA1c change, fasting glucose change) dominated the feature importance rankings. This confirms that including diagnostic markers creates circular reasoning, as these variables inherently define the disease outcome. Our main analysis deliberately excluded these diagnostic variables to assess the genuine predictive value of lifestyle and non-diagnostic biomarkers. Our analysis compared multiple machine learning approaches for predicting T2D onset across different feature domains and dimensionality reduction strategies. Table 2 presents the AUROC values for all model configurations. The ensemble model integrating predictions from all individual models generally delivered superior performance. The highest overall performance was achieved using all features with the hybrid approach (ensemble: 0.763, XGBoost: 0.752), followed by the original features approach (ensemble: 0.754, XGBoost: 0.738). Among feature domains, metabolic and biochemical markers showed the strongest predictive power (ensemble: 0.759 with original features), while lifestyle and socioeconomic factors alone demonstrated limited utility (AUROC < 0.6). To assess the robustness of our model selection, we performed additional 5-fold cross-validation analysis within the training set (80%) of the ASAS cohort. The cross-validation results showed consistent performance with our primary train-validation-test split approach, with AUROC differences of less than 0.03 across all models ( Supplementary Figure 3 ), confirming the stability of our model selection process. XGBoost consistently outperformed other individual algorithms across most feature combinations, with LGBM showing competitive results. The hybrid strategy (Original+PC) generally improved performance compared to original features alone, whereas exclusive use of PCA-based features resulted in reduced performance across all domains. Notably, combining lifestyle factors with anthropometric measurements (non-blood-based factors) substantially enhanced prediction capabilities compared to lifestyle factors alone, highlighting the synergistic effect of multi-domain feature integration. Table 3 highlights top model performance per domain. The ensemble model with hybrid features achieved the highest overall performance using all features (AUROC 0.763, F1 0.728, accuracy 0.680), closely followed by XGBoost (AUROC 0.752). Among feature domains, metabolic and biochemical markers demonstrated the strongest predictive capability, with the ensemble model using original features achieving AUROC 0.759 and F1 0.729. Anthropometric and physical measurements showed moderate performance, with the ensemble hybrid approach reaching AUROC 0.679 and notably high recall (0.909). Lifestyle and socioeconomic factors alone had limited predictive power, with the best model (LGBM hybrid) achieving AUROC 0.595. When combining non-blood-based factors, the ensemble model with original features achieved AUROC 0.686 and F1 0.711, demonstrating the synergistic effect of integrating lifestyle and anthropometric data. Table 2. Model Performance Comparison by Feature Domain and Dimensionality Reduction Strategy (AUROC) Feature Domain PCA Strategy LR RF XGBoost LGBM ANN RNN Ensemble All Features Original Features 0.723 0.740 0.738 0.74 2 0.637 0.691 0.75 4 Hybrid (Original+PC) 0.725 0.735 0.752 0.750 0.601 0.693 0.763 PC-Based 0.662 0.719 0.677 0.690 0.636 0.500 0.696 (a) Metabolic and biochemical markers Original Features 0.707 0.731 0.746 0.73 8 0.500 0.686 0.75 9 Hybrid (Original+PC) 0.710 0.710 0.736 0.744 0.594 0.724 0.747 PC-Based 0.632 0.637 0.641 0.657 0.656 0.639 0.671 (b) Anthropometric and physical measurements Original Features 0.657 0.671 0.677 0.67 7 0.611 0.588 0.67 9 Hybrid (Original+PC) 0.659 0.662 0.652 0.686 0.637 0.613 0.679 PC-Based 0.639 0.671 0.662 0.665 0.637 0.624 0.667 (c) Lifestyle and socioeconomic factors Original Features 0.578 0.574 0.568 0.58 9 0.531 0.555 0.58 0 Hybrid (Original+PC) 0.580 0.572 0.586 0.595 0.533 0.515 0.583 PC-Based 0.557 0.565 0.574 0.549 0.497 0.535 0.537 (b+c) Non-blood-based factors Original Features 0.659 0.669 0.675 0.68 6 0.592 0.570 0.68 6 Hybrid (Original+PC) 0.660 0.653 0.671 0.678 0.618 0.560 0.670 PC-Based 0.639 0.636 0.618 0.645 0.640 0.613 0.652 * Bold indicates highest AUROC per domain; all values based on test set performance. Table 3. Performance Metrics for Best Models in Each Domain Feature Domain Model Best Strategy AUROC Precision Recall F1 Score Accuracy All Features XGBoost Hybrid 0.752 0.621 0.887 0.730 0.675 LGBM Hybrid 0.750 0.640 0.843 0.728 0.687 Ensemble Hybrid 0.763 0.631 0.860 0.728 0.680 (a) Metabolic and biochemical markers XGBoost Original Features 0.746 0.610 0.881 0.721 0.661 LGBM Hybrid 0.744 0.622 0.885 0.731 0.676 Ensemble Original Features 0.759 0.634 0.859 0.729 0.683 (b) Anthropometric and physical measurements XGBoost Original Features 0.677 0.551 0.967 0.702 0.592 LGBM Hybrid 0.686 0.564 0.943 0.705 0.609 Ensemble Hybrid 0.679 0.574 0.909 0.704 0.620 (c) Lifestyle and socioeconomic factors XGBoost Hybrid 0.586 0.500 1.000 0.666 0.503 LGBM Hybrid 0.595 0.504 0.985 0.667 0.511 Ensemble Hybrid 0.583 0.497 1.000 0.664 0.497 (b+c) Non-blood-based factors XGBoost Original Features 0.675 0.570 0.927 0.706 0.617 LGBM Original Features 0.686 0.557 0.942 0.700 0.599 Ensemble Original Features 0.686 0.562 0.968 0.711 0.610 * Best strategy indicates the feature engineering approach that yielded the highest AUROC for each model within the respective domain. 3.3. Feature Importance Analysis Table 4 presents the top 10 feature importance rankings for LGBM and XGBoost models across different feature domains. LGBM prioritized TG as the most important predictor, followed by ▲CRP and ALT, while XGBoost ranked ▲CRP first, followed by ▲ALT and TG. This algorithmic difference demonstrates distinct preferences, with XGBoost showing stronger emphasis on change variables compared to LGBM's balance between static and dynamic measurements. Figure 2 provides SHAP value distributions for the hybrid model using all features, offering deeper insights into how individual features contribute to T2D predictions. The SHAP analysis reveals distinct patterns in feature contributions, where ▲CRP (crp_diff) and TG show the widest impact ranges and highest feature values, confirming their dominant predictive influence observed in the rankings. For key metabolic markers, higher feature values (red points) generally shifted toward positive SHAP values, indicating that elevated levels of ▲CRP, TG, ALT, and ▲ALT increase T2D risk. The clear separation between high and low feature values in the SHAP distributions validates the ranking patterns observed in Table 4 . The convergence of evidence from both feature importance rankings and SHAP impact analyses consistently highlights several key findings. Change variables demonstrated superior predictive value for inflammatory and liver markers (▲CRP, ▲ALT), while static TG levels (pre-diabetic values) showed greater importance than TG changes (▲TG) in both ranking positions and SHAP impact magnitudes. The metabolic markers ▲CRP and TG not only achieved top rankings but also showed the strongest predictive impacts across both analytical approaches. This multi-faceted analysis confirms that temporal changes in key biomarkers provide the most robust signals for T2D risk assessment, with the SHAP distributions providing mechanistic validation of the feature importance hierarchies. risk assessment. To further validate the clinical interpretability of our models, we conducted detailed tree structure analysis of the XGBoost model. Examination of decision nodes revealed clinically meaningful split thresholds, such as BMI < 25.56 and waist circumference < 88.9 cm, which align closely with established clinical criteria for metabolic risk assessment ( Supplementary Methods S2.4 , Supplementary Figure 4 ). This analysis confirmed that the model was learning biologically plausible decision criteria rather than merely partitioning data mechanically, supporting the clinical applicability of our predictive framework. Complete tree structures and decision paths are provided in the supplementary materials to enhance reproducibility and transparency. Table 4. Top 10 Feature Importance Rankings by Domain in LGBM and XGBoost Models Rank 1 2 3 4 5 6 7 8 9 10 All Features LGBM TG ▲CRP ALT WBC ▲TG Bio_ PCA_3 Physical_ PCA_2 ▲WBC ▲AST CRP XGBoost ▲CRP ▲ALT TG ALT Physical_ PCA_2 ▲TG ▲WBC ▲AST Bio_ PCA_2 ▲Hematocrit (a) Metabolic and biochemical markers LGBM TG ▲TG ▲CRP ▲WBC Age CRP ALT ▲Total_ cholesterol Bio_ PCA_10 Bio_ PCA_3 XGBoost ▲CRP TG ▲TG ▲Hemoglobin ▲ALT ALT ▲AST ▲WBC ▲RBC ▲Hematocrit (b) Anthropometric and physical measurements LGBM ▲Weight ▲BMI Weight ▲WC ▲DBP Physical_PCA_4 BMI WC Physical_ PCA_5 Physical_ PCA_3 XGBoost BMI ▲BMI ▲SBP ▲Weight SBP Residential_area WC ▲WC ▲DBP Sex (c) Lifestyle and socioeconomic factors LGBM Age ▲Medical_cost Life_ PCA_8 ▲Total_ cholesterol Life_PCA_10 Life_PCA_2 Residential_area Life_ PCA_4 Life_ PCA_6 ▲pack_ years XGBoost Smoking ▲Medical_cost ▲Smoking Residential_area Age ▲Urinary_frequency ▲pack_ years Sex Life_PCA_6 ▲Total_ cholesterol (b+c) Non-blood-based factors LGBM BMI ▲Weight WC ▲BMI ▲WC ▲SBP Age SBP Weight ▲DBP XGBoost ▲BMI BMI ▲SBP ▲Smoking ▲Weight ▲ WC Residential_area WC SBP ▲Urinary_frequency Abbreviations: BMI, body mass index; TG, triglycerides; CRP, C-reactive protein; ALT, alanine aminotransferase; BUN, blood urea nitrogen; WBC, white blood cell count; SBP, systolic blood pressure; DBP, diastolic blood pressure; WC, waist circumference; Hb, hemoglobin; Tchl, total cholesterol; packyr_current, pack-years for current smokers. Features marked with ▲ represent change values (difference between follow-up Vₙ₋₁-Vₙ) , indicating increases from the previous visit. 4. Discussion This study represents the first comprehensive investigation to develop machine learning models for predicting T2D onset by simultaneously incorporating both static and dynamic measurements of risk factors using population-based Korean cohort data. By employing an approach that excludes diagnostic glycemic markers while systematically considering longitudinal changes alongside static measurements in lifestyle factors and biomarkers, we present clinically relevant predictive models with significant translational potential. While prospective cohort studies analyzing T2D risk among healthy adults are extremely limited, our findings align with recent machine learning approaches for T2D prediction, with our ensemble model achieving an AUROC of 0.763, comparable to the Fasa Adult Cohort Study (FACS) by Talebi Moghaddam et al. (AUC 89.61% in 7,408 Iranian adults over 5 years) 23 . Consistent with our findings, the FACS study demonstrated superior performance of ensemble methods, with Random Forest-based approaches achieving optimal results, reinforcing the robustness of tree-based algorithms for T2D prediction in prospective cohort studies 5 . A key distinction is our simultaneous incorporation of both static baseline measurements and longitudinal changes of risk factors, whereas the FACS study relied primarily on static measurements despite its prospective design. This temporal approach allows our model to capture dynamic metabolic risk evolution over time, potentially explaining disease progression rather than merely identifying high-risk individuals at baseline. Various machine learning algorithms have been utilized in previous T2D prediction studies 4-6,18 . Systematic reviews of diabetes prediction models revealed that the vast majority of studies (>90%) employed cross-sectional or retrospective designs, with prospective cohort studies being extremely rare 5 . Tree-based algorithms, particularly XGBoost, Random Forest, and ensemble methods, consistently achieved the highest performance across studies 4-6,18 . Cross-sectional studies with small, well-curated datasets (such as Pima Indians Diabetes Dataset 24 ) achieved exceptional performance with AUC values of 0.85-0.95, largely attributed to the inclusion of diagnostic glycemic markers as predictive features, while studies using large-scale, real-world clinical datasets showed more modest performance with AUC ranges of 0.70-0.85 across diverse populations 4-6,18 . Notably, deep learning methods generally underperformed compared to tree-based algorithms despite their theoretical advantages for complex data patterns 25 . The significant performance gap between controlled datasets and real-world applications highlights the ongoing challenge of translating laboratory success to clinical practice. Domain-specific analysis revealed that metabolic and biochemical markers exhibited the most robust predictive power (Ensemble: AUROC 0.759), reflecting that T2D is fundamentally a metabolic disorder. Among these metabolic markers, change in C-reactive protein (▲CRP) emerged as the most important predictor in our feature importance analysis, highlighting the critical role of inflammatory processes in T2D pathogenesis. This finding aligns with previous longitudinal studies demonstrating that elevated CRP levels independently predict incident diabetes 26 , with a comprehensive meta-analysis of 22 cohorts involving 40,735 participants showing a 26% increased risk of T2D development in individuals with elevated CRP levels (RR 1.26; 95% CI 1.16-1.37) 27 . Notably, triglycerides (TG) ranked as the second most important predictor in LGBM models and third in XGBoost models, with static TG values demonstrating greater predictive importance than TG changes (▲TG), suggesting that baseline lipid levels may be more informative for diabetes risk assessment than dynamic lipid fluctuations. Both static BMI and BMI changes (▲BMI) showed significant predictive value, consistently ranking among the top predictors in non-blood-based factor analysis. Additionally, medical cost changes (▲Medical_cost) emerged as an important lifestyle predictor, potentially reflecting increased healthcare utilization patterns preceding T2D diagnosis and suggesting that healthcare usage patterns may serve as early warning signals for population screening strategies. Our methodological approach introduces three distinct innovations that address key limitations in current T2D prediction research. First, we systematically evaluated multi-domain feature integration, demonstrating that combining non-blood-based factors (lifestyle + anthropometric; AUROC 0.686) achieved moderate predictive performance while representing modifiable risk factors that can guide practical interventions. Although this performance is substantially lower than biochemical markers (AUROC 0.759), these non-blood-based factors offer significant clinical value for early prevention strategies and risk screening in resource-limited settings where blood testing is not readily available. Second, we applied domain-stratified PCA within biologically meaningful feature groups (lifestyle, anthropometric, biochemical) rather than global dimensionality reduction across all variables 12,14 , which preserves biological interpretability while effectively addressing multicollinearity. Third, unlike conventional static risk assessment approaches, our temporal change-based features capture dynamic metabolic transitions preceding T2D onset, enabling detection of disease progression patterns rather than baseline risk stratification This study has several important limitations. First, our study was conducted in a single population cohort in South Korea, which may limit the generalizability of our findings to other populations with different genetic backgrounds, healthcare systems, and T2D risk profiles. Additionally, external validation was not feasible because other cohorts lacked the comprehensive diabetes diagnostic measurements (FBG, HbA1c, and 2-hour blood glucose levels after a 75g OGTT) that were consistently available in the Ansan-Ansung cohort, preventing us from applying our models to independent datasets for validation. Second, despite our rigorous exclusion criteria and comprehensive data cleaning procedures, residual confounding from unmeasured variables cannot be completely ruled out. Third, while our longitudinal design captures important temporal changes in biomarkers, the 2-year follow-up period may not fully capture the long-term dynamics of diabetes development, and longer follow-up studies would be valuable. Fourth, although we excluded participants with diagnostic glycemic indicators to avoid circular reasoning, this approach may have inadvertently excluded individuals in the very early stages of glucose dysregulation who might benefit most from early intervention. Finally, our feature selection process, while systematic, was limited to variables available in the cohort database. Notably, dietary data, which are crucial for diabetes prediction, could not be included as dietary surveys were conducted only twice during the study period (baseline and third follow-up), providing insufficient temporal resolution for our longitudinal modeling approach. This limitation potentially missed important nutritional and lifestyle factors that could enhance prediction accuracy. In conclusion, dynamic changes in multiple risk factors over a 2-year period provided superior T2D predictive value compared to static measurements. The ensemble model achieved optimal performance (AUROC 0.763), closely followed by XGBoost (AUROC 0.752), with key predictors including inflammatory markers (▲CRP), lipid levels (TG), and anthropometric changes (▲BMI, ▲Weight). These findings offer significant potential for population-based screening, though external validation warrants further investigation. Declarations Declaration of competing interest The authors have no competing interests to report. Data availability The data supporting the findings of this study are available from the Korean Genome and Epidemiology Study (KoGES) Ansan-Ansung cohort. Restrictions apply to the availability of these data, which were used under license for the current study. Data access requests can be submitted through the Korea Disease Control and Prevention Agency (KDCA) at https://coda.nih.go.kr/usab/koges/intro.do. Source code for data analysis and visualization is publicly available at https://github.com/tgkim-kr/MRSi-AI. Author contributions Conceptualization: HWW. Methodology: TK, HWW. Formal analysis: TK, JD, HWW. Writing - original draft: TK, HWW. Writing - review and editing: TK, JD, HWW. Funding This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-RS-2023-00246789) References Cho, N. H. et al. IDF Diabetes Atlas: Global estimates of diabetes prevalence for 2017 and projections for 2045. Diabetes Res. Clin. Pract. 138 , 271–281. 10.1016/j.diabres.2018.02.023 (2018). Banday, M. Z., Sameer, A. S. & Nissar, S. Pathophysiology of diabetes: An overview. Avicenna J. Med. 10 , 174–188. 10.4103/ajm.ajm_53_20 (2020). Sarkar, S. et al. The onset and the development of cardiometabolic aging: an insight into the underlying mechanisms. Front. Pharmacol. 15 10.3389/fphar.2024.1447890 (2024). Nazirun, N. N. N. et al. Prediction Models for Type 2 Diabetes Progression: A Systematic Review. IEEE Access. 12 , 161595–161619. 10.1109/ACCESS.2024.3432118 (2024). Fregoso-Aparicio, L., Noguez, J., Montesinos, L. & García-García, J. A. Machine learning and deep learning predictive models for type 2 diabetes: a systematic review. Diabetol. Metab. Syndr. 13 , 148 (2021). Asgari, S., Khalili, D., Hosseinpanah, F. & Hadaegh, F. Prediction Models for Type 2 Diabetes Risk in the General Population: A Systematic Review of Observational Studies. Int. J. Endocrinol. Metab. 19 , e109206. 10.5812/ijem.109206 (2021). Naveed, I., Kaleem, M. F., Keshavjee, K. & Guergachi, A. Artificial intelligence with temporal features outperforms machine learning in predicting diabetes. PLOS Digit. Health . 2 , e0000354. 10.1371/journal.pdig.0000354 (2023). Lugner, M., Rawshani, A., Helleryd, E. & Eliasson, B. Identifying top ten predictors of type 2 diabetes through machine learning analysis of UK Biobank data. Sci. Rep. 14 , 2102. 10.1038/s41598-024-52023-5 (2024). Ejiyi, C. J. et al. A robust predictive diagnosis model for diabetes mellitus using Shapley-incorporated machine learning algorithms. Healthc. Analytics . 3 , 100166. https://doi.org/10.1016/j.health.2023.100166 (2023). Kim, Y. & Han, B. G. & group, t. K. Cohort Profile: The Korean Genome and Epidemiology Study (KoGES) Consortium. International Journal of Epidemiology 46, e20-e20, (2016). 10.1093/ije/dyv316 Committee, A. D. & A. P., P. Committee:, A. D. A. P. P. 2. Classification and diagnosis of diabetes: standards of medical care in diabetes—2022. Diabetes care . 45 , S17–S38 (2022). Nilashi, M. et al. Knowledge Discovery and Diseases Prediction: A Comparative Study of Machine Learning Techniques. Journal Soft Comput. & Decis. Support Systems 4 (2017). Guyon, I. & Elisseeff, A. An introduction to variable and feature selection. J. Mach. Learn. Res. 3 , 1157–1182 (2003). Changala, R. & Rao, D. R. Development of predictive model for medical domains to predict chronic diseases (diabetes) using machine learning algorithms and classification techniques. ARPN J. Eng. Appl. Sci. 14 , 1202–1212 (2019). Breiman, L. & Random Forests Mach. Learn. 45 , 5–32, doi: 10.1023/A:1010933404324 (2001). Chen, T. & Guestrin, C. in Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining. 785–794. Ke, G. et al. Lightgbm: A highly efficient gradient boosting decision tree. Advances neural Inform. Process. systems 30 (2017). Afsaneh, E., Sharifdini, A., Ghazzaghi, H. & Ghobadi, M. Z. Recent applications of machine learning and deep learning models in the prediction, diagnosis, and management of diabetes: a comprehensive review. Diabetol. Metab. Syndr. 14 , 196 (2022). Gardner, M. W. & Dorling, S. R. Artificial neural networks (the multilayer perceptron)—a review of applications in the atmospheric sciences. Atmos. Environ. 32 , 2627–2636 (1998). Mienye, I. D., Swart, T. G. & Obaido, G. Recurrent Neural Networks: A Comprehensive Review of Architectures, Variants, and Applications. Information 15 , 517 (2024). Oh, T. et al. Machine learning-based diagnosis and risk factor analysis of cardiocerebrovascular disease based on KNHANES. Sci. Rep. 12 , 2250. 10.1038/s41598-022-06333-1 (2022). Mujahid, M. et al. Data oversampling and imbalanced datasets: an investigation of performance for machine learning and feature engineering. J. Big Data . 11 , 87. 10.1186/s40537-024-00943-4 (2024). Talebi Moghaddam, M. et al. Predicting diabetes in adults: identifying important features in unbalanced data over a 5-year cohort study using machine learning algorithm. BMC Med. Res. Methodol. 24 , 220. 10.1186/s12874-024-02341-z (2024). Gupta, H., Varshney, H., Sharma, T. K., Pachauri, N. & Verma, O. P. Comparative performance analysis of quantum machine learning with deep learning for diabetes prediction. Complex. Intell. Syst. 8 , 3073–3087. 10.1007/s40747-021-00398-7 (2022). Zou, Q. et al. Predicting diabetes mellitus with machine learning techniques. Front. Genet. 9 , 515 (2018). Pradhan, A. D., Manson, J. E., Rifai, N., Buring, J. E. & Ridker, P. M. C-reactive protein, interleukin 6, and risk of developing type 2 diabetes mellitus. Jama 286 , 327–334. 10.1001/jama.286.3.327 (2001). Wang, X. et al. Inflammatory markers and risk of type 2 diabetes: a systematic review and meta-analysis. Diabetes Care . 36 , 166–175. 10.2337/dc12-0702 (2013). Maniruzzaman, M. et al. Accurate Diabetes Risk Stratification Using Machine Learning: Role of Missing Value and Outliers. J. Med. Syst. 42 , 92. 10.1007/s10916-018-0940-7 (2018). Additional Declarations No competing interests reported. Supplementary Files SupplementaryT2DASAS2510ff.docx Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 22 Dec, 2025 Reviewers agreed at journal 02 Dec, 2025 Reviewers agreed at journal 01 Dec, 2025 Reviewers invited by journal 01 Dec, 2025 Editor invited by journal 30 Oct, 2025 Editor assigned by journal 23 Oct, 2025 Submission checks completed at journal 23 Oct, 2025 First submitted to journal 20 Oct, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7902454","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":554331558,"identity":"bddcd5d8-231a-432e-bea8-5ddc4edc5141","order_by":0,"name":"Kim Taegyun","email":"","orcid":"","institution":"Medysapiens","correspondingAuthor":false,"prefix":"","firstName":"Kim","middleName":"","lastName":"Taegyun","suffix":""},{"id":554331559,"identity":"3ca68db7-1b49-47e0-9f34-9a2ab4b445eb","order_by":1,"name":"Jiwon Do","email":"","orcid":"","institution":"IT Support Office, Kyobo Book Centre","correspondingAuthor":false,"prefix":"","firstName":"Jiwon","middleName":"","lastName":"Do","suffix":""},{"id":554331560,"identity":"be17e4f7-b644-4ee7-ad39-7ad8c91ab1aa","order_by":2,"name":"Hye Won Woo","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA5ElEQVRIiWNgGAWjYHACNgbGBgY5AyQRZqK0GJOuJXED0VoMjjcfe/Bxh036dvazxyR+7qhl4G8/wGxcgU/LmWPphjPPpOXu7MlLk+w9c5xB4kwCc+IZPFrMbuSYSfO2Hc7dcCDHTIK37RgDww0G5oMNhLT8bfufbnD+jZnkX6AWeaK0MLYdSDCAWFfDYADUkohPi/2ZY0AvtCUbbrjxxthatu0Aj+GZxGZDfFok25uBAdVmJ29wPsfw5tu2Ojm544cPS+LTggxYJBgYDvMwgKKJWMD8gYGhjmjVo2AUjIJRMHIAAK7KUK0cMzQRAAAAAElFTkSuQmCC","orcid":"","institution":"Hanyang University","correspondingAuthor":true,"prefix":"","firstName":"Hye","middleName":"Won","lastName":"Woo","suffix":""}],"badges":[],"createdAt":"2025-10-20 05:53:20","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7902454/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7902454/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":97487532,"identity":"71321831-57f1-473a-8dc5-288bba015a35","added_by":"auto","created_at":"2025-12-05 01:50:50","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1584019,"visible":true,"origin":"","legend":"","description":"","filename":"ManuscriptT2DASAS251022f.docx","url":"https://assets-eu.researchsquare.com/files/rs-7902454/v1/618c3711e61cdefce070235a.docx"},{"id":97487535,"identity":"9af98ad3-84aa-4aae-80c4-ae6103be02f5","added_by":"auto","created_at":"2025-12-05 01:50:50","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6191,"visible":true,"origin":"","legend":"","description":"","filename":"c6c574d5e7904aa1bf99de5dc01b6054.json","url":"https://assets-eu.researchsquare.com/files/rs-7902454/v1/416a403ab6ceeda1c90d3507.json"},{"id":97487540,"identity":"ea0d8d83-0ce1-4ede-a1f6-94a6adcbd414","added_by":"auto","created_at":"2025-12-05 01:50:50","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1522202,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryT2DASAS2510ff.docx","url":"https://assets-eu.researchsquare.com/files/rs-7902454/v1/c1117955c989a1f20caf4aad.docx"},{"id":97487538,"identity":"a7dae2a2-1a12-4e59-ab62-fe63efec87f8","added_by":"auto","created_at":"2025-12-05 01:50:50","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":143141,"visible":true,"origin":"","legend":"","description":"","filename":"c6c574d5e7904aa1bf99de5dc01b60541enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7902454/v1/f7aa2119c282da9f995f1d13.xml"},{"id":97669882,"identity":"e5366655-343d-404b-8cc2-061c3a247e4f","added_by":"auto","created_at":"2025-12-08 09:29:16","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":230415,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7902454/v1/0f5919b5065d84d622c29065.png"},{"id":97670476,"identity":"0ec03c2a-6c3d-4cda-ba1b-97306a718cfe","added_by":"auto","created_at":"2025-12-08 09:30:47","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":98577,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7902454/v1/b08ab42cc259b311d5cac254.png"},{"id":97487542,"identity":"102b2680-ba66-4d7d-b764-d38b61c82ef3","added_by":"auto","created_at":"2025-12-05 01:50:51","extension":"xml","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":144319,"visible":true,"origin":"","legend":"","description":"","filename":"c6c574d5e7904aa1bf99de5dc01b60541structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7902454/v1/cfb44abcdc8de72d5f82e9d6.xml"},{"id":97487541,"identity":"7ed47b20-3f8b-44f0-bb0c-c71e55cfd5d3","added_by":"auto","created_at":"2025-12-05 01:50:50","extension":"html","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":153374,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7902454/v1/465bdd556ef6dd5b953f38f3.html"},{"id":97669469,"identity":"feb9dcfa-f809-41a7-832f-48b0a14fe45f","added_by":"auto","created_at":"2025-12-08 09:28:04","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":863203,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverview of the machine learning pipeline for Type 2 diabetes prediction\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7902454/v1/da9ade931d510d5844cf2bc3.png"},{"id":97670629,"identity":"727eb320-efd1-44d9-b78a-9b22c87d8a5e","added_by":"auto","created_at":"2025-12-08 09:31:03","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":152470,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eHAP Feature Importance Summary Plots for All Features (Hybrid) Models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSHAP (SHapley Additive exPlanations) summary plots showing feature importance and impact on model predictions for (A) LGBM and (B) XGBoost models. Change variables (marked with ▲) represent the difference between current visit (Vₙ) and previous visit (Vₙ₋₁), calculated as Vₙ - Vₙ₋₁. Key abbreviations include: crp_diff (▲CRP, change in high-sensitivity C-reactive protein), tg (triglycerides), alt (alanine aminotransferase), alt_diff (▲ALT, change in alanine aminotransferase), physical_pca_2 (physical component analysis factor 2), wbc_diff (▲WBC, change in white blood cell count), bmi (body mass index), sbp (systolic blood pressure), age (age in years), and as1_area (residential area: Ansung=1, Ansan=0).\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7902454/v1/78279b81ad3fa749129915a8.jpg"},{"id":97677794,"identity":"573c3fce-7c23-4919-8d4a-ba6fae75ce17","added_by":"auto","created_at":"2025-12-08 09:54:33","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2504811,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7902454/v1/3e469ccb-a07f-430d-adff-f7500257c0f2.pdf"},{"id":97487537,"identity":"1f2c7a30-3710-4715-92a1-f7b0e65da3af","added_by":"auto","created_at":"2025-12-05 01:50:50","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":1522202,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryT2DASAS2510ff.docx","url":"https://assets-eu.researchsquare.com/files/rs-7902454/v1/4bdfaf946992ff0b72340b76.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Predicting Type 2 Diabetes Using Baseline and Longitudinal Changes in Lifestyle and Clinical Markers: A Machine Learning Approach","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eType 2 Diabetes (T2D) remains a global health burden, affecting 451\u0026nbsp;million people worldwide in 2017 and projected to reach 693\u0026nbsp;million by 2045 \u003csup\u003e1\u003c/sup\u003e. The development of T2D is typically characterized by either a gradual accumulation of risk factors from middle adulthood, a sudden deterioration in health shortly before diagnosis, or a combination of both trajectories \u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. In recent years, there has been growing interest in leveraging machine learning techniques to improve the prediction of T2D onset \u003csup\u003e\u003cspan additionalcitationids=\"CR5\" citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eCurrent machine learning approaches for T2D prediction face two critical limitations. First, these models predominantly rely on cross-sectional or static baseline data, failing to capture the temporal dynamics of risk factor evolution\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. This is problematic because diabetes risk accumulates over time\u0026mdash;an individual with elevated risk factors for 10 years has substantially different risk than someone with identical readings for one year, yet most models cannot distinguish between these scenarios\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. Second, many existing models incorporate diagnostic glycemic markers (glucose, HbA1c) as input features\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e,\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e, creating a logical flaw of using disease-defining variables to predict the disease itself. These limitations collectively undermine model validity: current approaches neither capture temporal risk accumulation nor distinguish between genuine prediction and early disease detection. While machine learning shows promise in healthcare, there remains a critical gap in understanding how temporal changes in lifestyle factors and biological markers contribute to diabetes risk progression.\u003c/p\u003e\u003cp\u003eTo address these limitations, this study applies multiple machine learning approaches to develop prospective prediction models for T2D onset using both static measurements and interval changes between consecutive time points in lifestyle factors and biological markers. Using data from the Ansan and Ansung cohort of the Korean Genome and Epidemiology Study (KoGES), we exclude changes in blood glucose and HbA1c to avoid circular reasoning. To ensure transparency in clinical decision support, we apply explainable AI methods to interpret how individual features contribute to risk predictions.\u003c/p\u003e"},{"header":"2. Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1. Study population\u003c/h2\u003e\u003cp\u003eData were obtained from the Ansan\u0026ndash;Ansung cohort of the Korean Genome and Epidemiology Study (KoGES), a 10-wave longitudinal cohort study of Korean adults aged 40 years or older at baseline (2001\u0026ndash;2002), with biannual follow-ups conducted through 2019\u0026ndash;2020. The Ansan\u0026ndash;Ansung study is one of the KoGES population-based cohort studies involving community-dwelling individuals who participated in health examinations. It comprises men and women aged 40\u0026ndash;69 years at baseline, residing in Ansan (an urban area) and Ansung (a rural area). All participants voluntarily participated and provided written informed consent. The study was conducted in accordance with the Declaration of Helsinki and was approved by the ethics committee of the Korean Health and Genome Study, supported by the Korea National Institute of Health (KNIH). Further details on the KoGES design and sampling methods have been described in previous reports \u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eFor the purpose of this prospective study, among the 10,038 participants in the baseline survey, we excluded 1,352 (13.5%) participants who met the American Diabetes Association (ADA) criteria for T2D \u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e: use of anti-diabetic medications or insulin, fasting blood glucose (FBG)\u0026thinsp;\u0026ge;\u0026thinsp;126 mg/dL (7.0 mmol/L), 2-hour blood glucose\u0026thinsp;\u0026ge;\u0026thinsp;200 mg/dL (11.1 mmol/L) during an oral glucose tolerance test (OGTT), or glycated hemoglobin (HbA1c)\u0026thinsp;\u0026ge;\u0026thinsp;6.5%. Furthermore, we excluded those with a history of cardiovascular disease, cerebrovascular disease, or cancer at baseline (n\u0026thinsp;=\u0026thinsp;439) to eliminate potential confounding effects from these more severe conditions that could alter lifestyle behaviors and medication use. We also excluded participants who did not attend follow-up visits (n\u0026thinsp;=\u0026thinsp;737). As a result, 7,510 participants (3,565 men and 3,945 women) were included in the final analysis.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2. Ascertainment of diabetes incidence\u003c/h2\u003e\u003cp\u003eParticipants underwent nine biennial follow-up examinations from 2003\u0026ndash;2004 to 2019\u0026ndash;2020. At each visit, FBG, HbA1c, and 2-hour blood glucose levels after a 75g OGTT were measured. Using the same ADA criteria applied at baseline, incident T2D was assessed at each follow-up visit \u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. The detailed criteria for T2D identification are described in section \u003cspan refid=\"Sec7\" class=\"InternalRef\"\u003e2.4.1\u003c/span\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3. Assessment of lifestyle factors and biomarkers\u003c/h2\u003e\u003cp\u003eAll predictor variables were collected at each follow-up visit by trained interviewers following a standardized protocol. Interviewers were trained to administer the structured questionnaire based on a standardized manual \u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. Predictor variables were categorized into three domains: lifestyle and socioeconomic factors, anthropometric and physical measurements, and metabolic and biochemical markers.\u003c/p\u003e\u003cp\u003eLifestyle and socioeconomic factors included smoking status (current, former, or never), pack-years (for current smokers), alcohol drinking status and alcohol consumption (categorized as current, former, or never; total alcohol intake was assessed only among current drinkers), physical activity (defined as moderate-intensity activity\u0026thinsp;\u0026ge;\u0026thinsp;3 times/week, \u0026ge;\u0026thinsp;30 minutes/session), water intake frequency, urinary frequency, residential area (Ansan or Ansung), education level (high school graduate: yes or no), parental history of diabetes (yes or no), and housing type (no housing, owner-occupied, deposit-based or monthly rental, and other). Economic indicators included monthly average income, medical expenses, and expenditures on health foods.\u003c/p\u003e\u003cp\u003eAnthropometric and physical measurements were obtained by trained health professionals using standardized procedures. Body weight was measured to the nearest 0.1 kg with participants wearing light indoor clothing without shoes. Height was measured to the nearest 0.1 cm. Body mass index (BMI) was calculated as weight (kg) divided by height squared (m\u0026sup2;). Waist circumference (WC) was measured at the narrowest part between the lower rib and the iliac crest to the nearest 0.1 cm. Blood pressure was measured three times in the right upper arm in the supine position using a standard mercury sphygmomanometer, and the average of the second and third readings was used for analysis.\u003c/p\u003e\u003cp\u003eMetabolic and biochemical markers were assessed from blood samples collected after overnight fasting. All samples were analyzed at a central laboratory (Seoul Clinical Laboratories, Seoul, Korea). Serum glucose, triglycerides, total cholesterol, HDL-cholesterol, liver enzymes (AST, ALT), kidney function markers (blood urea nitrogen [BUN], serum creatinine), albumin, and inflammatory marker (high-sensitivity C-reactive protein [hsCRP]) were measured using an automated chemistry analyzer (ADVIA 1650; Siemens, NY, USA). Complete blood count was used to determine white blood cell count (WBC), red blood cell count (RBC), platelet count (PLT), hematocrit (Hct), and hemoglobin (Hb). For measurements obtained prior to the introduction of ADVIA 1650 (September 2002), validated conversion equations were applied to ensure compatibility across instruments and consistency over time.\u003c/p\u003e\u003cp\u003eAmong all predictor variables, sex, residential area, education level, parental history of diabetes, and housing type were considered fixed baseline predictors. Missing values in the dataset were handled through imputation, where continuous variables were imputed using mean values and categorical variables were imputed using mode values to maintain data completeness for subsequent machine learning analysis.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.4. Feature engineering and Model Development\u003c/h2\u003e\u003cdiv id=\"Sec7\" class=\"Section3\"\u003e\u003ch2\u003e2.4.1 Target Variable Definition\u003c/h2\u003e\u003cp\u003eTo enhance predictive precision and account for potential confounding effects of medication on clinical parameters, we defined the target outcome using a laboratory parameter-based definition exclusively. We established T2D onset identification solely through objective measurements (FBG\u0026thinsp;\u0026ge;\u0026thinsp;126 mg/dL, 2-hour plasma glucose\u0026thinsp;\u0026ge;\u0026thinsp;200 mg/dL after OGTT, or HbA1c\u0026thinsp;\u0026ge;\u0026thinsp;6.5%) without including medication-based diagnoses. This approach enabled analysis of natural disease biomarker progression without pharmacological confounding, as anti-diabetic medications can normalize glucose levels, alter other clinical parameters, and prompt behavioral changes.\u003c/p\u003e\u003cp\u003eFor participants who initiated anti-diabetic medications or insulin therapy during the follow-up period, we censored their data at the last visit where they remained medication-free and maintained non-diabetic status. This censoring approach ensured that our analysis captured only the natural disease progression trajectory, avoiding potential bias introduced by treatment effects on parameters.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec8\" class=\"Section3\"\u003e\u003ch2\u003e2.4.2 Change-Based Feature Engineering\u003c/h2\u003e\u003cp\u003eTo capture temporal patterns in T2D onset among initially non-diabetic individuals, we utilized both static measurements (observed values at each visit) and generated change-based features representing interval changes in variables between consecutive visits. For participants who developed T2D, if T2D was diagnosed at visit Vₙ, the interval change (Vₙ-Vₙ₋₁) was considered indicative of the transition to diabetes, while all previous interval changes (V₂-V₁, V₃-V₂, ..., Vₙ₋₁-Vₙ₋₂) from the same participant represented normal variation. For example, if a participant was diagnosed at V₄, the change V₄-V₃ would represent diabetes-related transition, while changes V₂-V₁ and V₃-V₂ would represent normal fluctuations.\u003c/p\u003e\u003cp\u003eThis approach yielded 1,796 data points corresponding to visits immediately preceding diabetes diagnosis and 45,849 data points corresponding to pre-diagnostic visits from participants who remained non-diabetic throughout follow-up. The model was trained to predict the probability of T2D onset at visit Vₜ based on observed features at visit Vₜ₋₁ and their changes between Vₜ₋₁ and Vₜ. This approach enabled the model to identify patterns from both current visit measurements and their interval changes that distinguish imminent diabetes onset from normal fluctuation.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section3\"\u003e\u003ch2\u003e2.4.3 Dimensionality Reduction Using Principal Component Analysis (PCA)\u003c/h2\u003e\u003cp\u003eMost ML-based studies overlook the effect of multicollinearity on risk factor analysis \u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. Although multicollinearity does not affect the predictive power of ML models, variables with high collinearity offset the importance of each other, consequently leading to erroneous evaluation of variable importance \u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. Correlation analysis of our dataset revealed varying degrees of multicollinearity across different variable domains (\u003cb\u003eSupplementary Fig.\u0026nbsp;1\u003c/b\u003e), with metabolic and biochemical markers showing the strongest intercorrelations, followed by anthropometric measurements and lifestyle factors.\u003c/p\u003e\u003cp\u003eTo address multicollinearity issues, we applied Principal Component Analysis (PCA) transformation, which converts correlated input variables into orthogonal components. The input features were categorized into three biologically relevant domains: (1) lifestyle and socioeconomic factors, (2) anthropometric and physical measurements, and (3) metabolic and biochemical markers. PCA was applied separately within each domain to generate interpretable principal components while preserving domain-specific biological relationships.\u003c/p\u003e\u003cp\u003eWe employed three distinct analytical strategies to evaluate the optimal feature representation: (1) Original Features Analysis - using only the original variables without dimensionality reduction; (2) Hybrid Analysis - combining original variables with PCA-derived components; and (3) PC-Based Analysis - using exclusively the extracted principal components as predictors. This comparative approach enabled us to assess the explanatory power of transformed components versus their additive value when combined with original variables. Additionally, we evaluated model performance using domain-specific feature sets (lifestyle, anthropometric, biochemical) as well as combined non-blood-based factors (lifestyle\u0026thinsp;+\u0026thinsp;anthropometric) to identify the most informative feature combinations.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section3\"\u003e\u003ch2\u003e2.4.4 Model Development\u003c/h2\u003e\u003cdiv id=\"Sec11\" class=\"Section4\"\u003e\u003ch2\u003e2.4.4.1 Baseline Model\u003c/h2\u003e\u003cp\u003eWe employed logistic regression (LR) as the baseline model to establish a performance benchmark and assess the linear predictability of T2D onset. Logistic regression was selected for its interpretability and established use in clinical risk prediction, following the standard formulation:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:Logit\\left(p\\right)=\\:{\\beta\\:}_{0}+\\:\\sum\\:_{i=1}^{n}{\\beta\\:}_{i\\:}{x}_{i}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThis baseline approach enabled objective evaluation of whether improvements observed in more complex models constituted meaningful enhancements, while providing interpretable coefficients with corresponding statistical significance for each predictor.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section4\"\u003e\u003ch2\u003e2.4.4.2 Advanced Machine Learning Models\u003c/h2\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e illustrates the comprehensive machine learning pipeline developed for T2D prediction. The framework consists of three main phases: learning, validation, and inference. In the learning phase, pre-processed features from the ASAS cohort undergo model training using six different algorithms, which are then combined into an ensemble model.\u003c/p\u003e\u003cp\u003eWe implemented five machine learning algorithms to capture complex, non-linear relationships in T2D prediction: Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Artificial Neural Network (ANN), and Recurrent Neural Network (RNN). Detailed model architectures, hyperparameter settings, and training procedures are provided in \u003cb\u003eSupplementary Methods\u003c/b\u003e. Tree-based models (RF, XGBoost, LightGBM) were selected for their ability to capture non-linear relationships and feature interactions through recursive data partitioning \u003csup\u003e\u003cspan additionalcitationids=\"CR16\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. These models construct decision trees by selecting optimal splits based on impurity reduction criteria, effectively handling categorical variables and missing values without requiring feature scaling \u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. Neural network models (ANN, RNN) were implemented to learn complex non-linear patterns through weighted transformations across multiple layers. ANN captures high-dimensional feature interactions through hidden layers \u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e, while RNN was specifically designed to process sequential data by maintaining temporal dependencies through recurrent connections, enabling the model to capture patterns across time steps \u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThis framework illustrates the complete pipeline from model training using Ansan\u0026ndash;Ansung cohort data (Learning) to inference and generalization on population-level data (Inference). Six machine learning algorithms (logistic regression, random forest, XGBoost, LightGBM, RNN, and ANN) were integrated into an ensemble model and applied through the Multi-Domain Simulation Interface (MDSi) framework to assess diabetes risk across lifestyle, anthropometric, and metabolic domains.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section4\"\u003e\u003ch2\u003e2.4.4.3 Ensemble Model Implementation\u003c/h2\u003e\u003cp\u003eWe implemented a soft voting ensemble that combines predictions from all individual models by averaging their predicted probabilities. This approach integrates diverse model strengths while mitigating individual model limitations through complementary effects, particularly beneficial given the different architectural approaches and feature processing mechanisms of the base models.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003e2.5. Model Evaluation, Interpretability, and Feature Importance\u003c/h2\u003e\u003cp\u003eThe dataset exhibited severe class imbalance, with positive cases (T2D onset) representing only 3.9% of total instances. To address this imbalance, we applied post-split over-sampling with replacement within each dataset partition\u0026mdash;training (80%), validation (10%), and test (10%)\u0026mdash;after data splitting. Class imbalance is common in epidemiological data, making accurate performance assessment challenging without balancing technique\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. Over-sampling of the minority class was performed independently within each subset to achieve a balanced 1:1 class ratio while preventing information leakage and overfitting risk associated with pre-split oversampling\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eModel performance was evaluated using accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUROC) to comprehensively assess both overall and minority class performance. To assess model interpretability and feature importance, we employed model-specific approaches: regression coefficients and p-values for Logistic Regression, feature importance values for tree-based models (RF, XGBoost, LightGBM), and predictive performance evaluation for neural network models (ANN, RNN). The ensemble approach combined predictions from all individual models through soft voting to enhance overall predictive accuracy. Detailed procedures are provided in the \u003cb\u003eSupplementary Methods\u003c/b\u003e.\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Results","content":"\u003cp\u003e\u003cstrong\u003e3.1. Characteristics of the study participants\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAs shown in \u003cstrong\u003eTable 1\u003c/strong\u003e, a total of 7,510 participants were included in the final analysis, with 5,505 (73.3%) remaining non-diabetic and 2,005 (26.7%) developing incident T2D during follow-up. Among those who developed T2D, 1,796 (89.6%) were screen-detected cases identified through follow-up laboratory testing and 209 (10.4%) were medication-detected cases (for whom only pre-diabetic health data were included in the model).\u0026nbsp;The study population consisted of middle-aged Korean adults (mean age 51.5 \u0026plusmn; 8.7 years) with balanced sex distribution and equal urban-rural representation. \u003cstrong\u003eLifestyle and socioeconomic factors\u003c/strong\u003e revealed that future T2D cases had higher medical expenses, with increased prevalence of family diabetes history (9.4% vs. 7.1%) and current smoking (28.1% vs. 24.6%).\u0026nbsp;\u003cstrong\u003eAnthropometric and physical measurements\u003c/strong\u003e demonstrated significantly elevated baseline values in the T2D group, including higher BMI (25.1 \u0026plusmn; 3.1 vs. 24.1 \u0026plusmn; 3.0 kg/m\u0026sup2;), waist circumference (84.9 \u0026plusmn; 8.5 vs. 81.1 \u0026plusmn; 8.6 cm), and blood pressure (123.7/80.9 vs. 118.9/79.0 mmHg).\u0026nbsp;\u003cstrong\u003eMetabolic and biochemical markers\u0026nbsp;\u003c/strong\u003eshowed a distinct pre-diabetic profile in future T2D cases, with elevated glucose parameters even within normal ranges (FBG: 86.3 vs. 81.4 mg/dL; HbA1c: 5.7 vs. 5.48%), adverse lipid profiles (triglycerides: 180.2 vs. 145.2 mg/dL; HDL-cholesterol: 43.1 vs. 45.6 mg/dL), and higher inflammatory markers (hsCRP: 0.24 vs. 0.21 mg/dL), indicating subtle baseline metabolic differences that preceded future diabetes development, despite all participants being non-diabetic at study entry.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1. Baseline characteristics of the study\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"699\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eBaseline variables\u003csup\u003e*\u003c/sup\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eTotal\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eConsistently non-diabetic individuals\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eT2D cases detected through follow-up testing\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eT2D cases identified through medication initiation\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eParticipants, N\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cem\u003e7,510\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cem\u003e5,505\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cem\u003e1,796\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cem\u003e209\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eLifestyle and socioeconomic factors\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eResidential area (Ansung, %)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e51.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e52.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e46.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e56.0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eMale, %\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e47.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e46.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e41.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e51.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eAge, years\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e51.5 \u0026plusmn; 8.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e51.2 \u0026plusmn; 8.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e52.1 \u0026plusmn; 5.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e54.2 \u0026plusmn; 8.7\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eHigh school graduate\u003csup\u003e1\u003c/sup\u003e, %\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e13.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e13.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e14.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e8.7\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eType of home ownership (Own, %)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e83.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e82.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e83.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e82.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eAverage monthly income, ten thousand won\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e1,825 \u0026plusmn; 1,401\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e1,838 \u0026plusmn; 1,401\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e1,822 \u0026plusmn; 1,393\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e1,519 \u0026plusmn; 1,411\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eMedical expenses, ten thousand won\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e42.0 \u0026plusmn; 57.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e41.1 \u0026plusmn; 56.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e43.9 \u0026plusmn; 60.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e49.3 \u0026plusmn; 65.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eExpenditures on health foods, ten thousand won\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e31.1 \u0026plusmn; 56.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e31.4 \u0026plusmn; 57.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e29.9 \u0026plusmn; 54.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e35.7 \u0026plusmn; 64.0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eFamily history of diabetes in parents, %\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e7.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e7.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e9.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e8.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eRegular exercise\u003csup\u003e2\u003c/sup\u003e, %\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e59.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e60.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e58.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e53.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eCurrent smoker, %\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e25.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e24.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e28.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e26.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003ePack-years (for current smokers), y\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e6.0 \u0026plusmn; 13.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5.7 \u0026plusmn; 12.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e6.9 \u0026plusmn; 13.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e6.7 \u0026plusmn; 14.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eCurrent drinker, %\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e48.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e47.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e51.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e40.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eAlcohol consumption, g/d\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e9.2 \u0026plusmn; 21.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e8.9 \u0026plusmn; 20.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e10.6 \u0026plusmn; 22.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e7.5 \u0026plusmn; 21.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eRecent increase in frequency of drinking water, %\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e13.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e13.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e12.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e15.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eRecent increase in urination frequency, %\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e19.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e18.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e19.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e25.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eAnthropometric and physical measurements\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eBody weight, kg\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e62.7 \u0026plusmn; 10.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e62.0 \u0026plusmn; 9.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e64.7 \u0026plusmn; 10.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e65.5 \u0026plusmn; 10.7\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eWaist circumference, cm\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e82.1 \u0026plusmn; 8.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e81.3 \u0026plusmn; 8.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e84.9 \u0026plusmn; 8.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e84.2 \u0026plusmn; 8.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eBody Mass Index, kg/m\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e24.5 \u0026plusmn; 3.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e24.1 \u0026plusmn; 3.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e25.1 \u0026plusmn; 3.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e25.9 \u0026plusmn; 3.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eSystolic blood pressure (SBP), mmHg.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e120.3 \u0026plusmn; 17.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e118.9 \u0026plusmn; 17.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e123.7 \u0026plusmn; 18.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e127.6 \u0026plusmn; 18.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eDiastolic Blood Pressure (DBP), mmHg.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e79.9 \u0026plusmn; 11.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e79.0 \u0026plusmn; 11.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e80.9 \u0026plusmn; 11.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e84.4 \u0026plusmn; 11.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eMetabolic and biochemical markers\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eFasting blood glucose (FBG), mg/dL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e82.7 \u0026plusmn; 8.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e81.4 \u0026plusmn; 7.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e86.3 \u0026plusmn; 9.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e87.0 \u0026plusmn; 9.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eA haemoglobin A1C (HbA1C), \u003cstrong\u003e%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5.54\u0026nbsp;\u0026plusmn; 0.35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5.48 \u0026plusmn; 0.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5.7 \u0026plusmn; 0.34\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5.8 \u0026plusmn; 0.34\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eTriglycerides (TG), mg/dL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e155.0 \u0026plusmn; 95.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e145.2 \u0026plusmn; 87.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e180.2 \u0026plusmn; 102.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e198.9 \u0026plusmn; 160.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eTotal cholesterol (Tchl), mg/dL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e189.3 \u0026plusmn; 34.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e187.7 \u0026plusmn; 34.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e193.2 \u0026plusmn; 33.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e197.1 \u0026plusmn; 33.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eHDL-cholesterol, mg/dL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e44.9 \u0026plusmn; 10.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e45.6 \u0026plusmn; 10.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e43.1 \u0026plusmn; 9.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e42.7 \u0026plusmn; 10.6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eAspartate transferase (AST), U/L\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e29.2 \u0026plusmn; 16.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e28.6 \u0026plusmn; 16.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e30.8 \u0026plusmn; 18.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e29.0 \u0026plusmn; 9.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eAlanine aminotransferase (ALT), U/L\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e27.1 \u0026plusmn; 22.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e25.6 \u0026plusmn; 19.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e31.3 \u0026plusmn; 28.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e29.9 \u0026plusmn; 16.6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eBlood urea nitrogen (BUN), mg/dL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e14.3 \u0026plusmn; 3.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e14.2 \u0026plusmn; 3.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e14.5 \u0026plusmn; 3.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e14.6 \u0026plusmn; 3.6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eSerum creatinine, mg/dL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.84 \u0026plusmn; 0.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.83 \u0026plusmn; 0.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.85 \u0026plusmn; 0.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.82 \u0026plusmn; 0.16\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eHigh-sensitivity C-reactive protein (hsCRP), mg/dL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.22 \u0026plusmn; 0.52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.21 \u0026plusmn; 0.52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.24 \u0026plusmn; 0.48\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.32 \u0026plusmn; 0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eWhite blood cell count (WBC),\u0026nbsp;10\u003csup\u003e3\u003c/sup\u003e/\u0026micro;L\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e6.48 \u0026plusmn; 1.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e6.38 \u0026plusmn; 1.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e6.7 \u0026plusmn; 1.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e6.9 \u0026plusmn; 1.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eRed blood cell count (RBC),\u0026nbsp;10\u003csup\u003e3\u003c/sup\u003e/\u0026micro;L\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.41 \u0026plusmn; 0.46\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.39 \u0026plusmn; 0.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.48 \u0026plusmn; 0.42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.42 \u0026plusmn; 0.48\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003ePlatelet count (PLT),\u0026nbsp;10\u003csup\u003e3\u003c/sup\u003e/\u0026micro;L\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e265.9 \u0026plusmn; 62.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e264.5 \u0026plusmn; 62.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e269.6 \u0026plusmn; 64.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e271.3 \u0026plusmn; 66.0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eHematocrit (Hct),\u0026nbsp;%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e40.9 \u0026plusmn; 4.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e40.8 \u0026plusmn; 4.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e41.5 \u0026plusmn; 4.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e41.1 \u0026plusmn; 4.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eHemoglobin (Hb),\u0026nbsp;g/dL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e13.6 \u0026plusmn; 1.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e13.5 \u0026plusmn; 1.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e13.8 \u0026plusmn; 1.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e13.7 \u0026plusmn; 1.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e* The values are expressed as mean \u0026plusmn; SD for continuous variables or percentage for categorical variables. \u003csup\u003e\u003cbr\u003e\u0026nbsp;1\u003c/sup\u003e High school graduate (\u0026ge; 12 years of education).\u003c/p\u003e\n\u003cp\u003e\u003csup\u003e2\u003c/sup\u003e Regular exercise (\u0026ge; 3 times/week and \u0026ge; 30 min/session).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2. Model Performance Comparison\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo demonstrate the circular nature of including diagnostic glycemic markers, supplementary analysis showed AUROC values exceeding 0.9 \u003cstrong\u003eSupplementary Table 1\u003c/strong\u003e when blood glucose and HbA1c were included as predictors. \u003cstrong\u003eSupplementary Figure 2\u003c/strong\u003e illustrates the SHAP analysis revealing that glucose-related variables (change in 2-hour OGTT glucose, 2-hour OGTT glucose level, HbA1c change, fasting glucose change) dominated the feature importance rankings. This confirms that including diagnostic markers creates circular reasoning, as these variables inherently define the disease outcome. Our main analysis deliberately excluded these diagnostic variables to assess the genuine predictive value of lifestyle and non-diagnostic biomarkers.\u003c/p\u003e\n\u003cp\u003eOur analysis compared multiple machine learning approaches for predicting T2D onset across different feature domains and dimensionality reduction strategies. \u003cstrong\u003eTable 2\u003c/strong\u003e presents the AUROC values for all model configurations. The ensemble model integrating predictions from all individual models generally delivered superior performance. The highest overall performance was achieved using all features with the hybrid approach (ensemble: 0.763, XGBoost: 0.752), followed by the original features approach (ensemble: 0.754, XGBoost: 0.738). Among feature domains, metabolic and biochemical markers showed the strongest predictive power (ensemble: 0.759 with original features), while lifestyle and socioeconomic factors alone demonstrated limited utility (AUROC \u0026lt; 0.6). To assess the robustness of our model selection, we performed additional 5-fold cross-validation analysis within the training set (80%) of the ASAS cohort. The cross-validation results showed consistent performance with our primary train-validation-test split approach, with AUROC differences of less than 0.03 across all models (\u003cstrong\u003eSupplementary Figure 3\u003c/strong\u003e), confirming the stability of our model selection process.\u003c/p\u003e\n\u003cp\u003eXGBoost consistently outperformed other individual algorithms across most feature combinations, with LGBM showing competitive results. The hybrid strategy (Original+PC) generally improved performance compared to original features alone, whereas exclusive use of PCA-based features resulted in reduced performance across all domains. Notably, combining lifestyle factors with anthropometric measurements (non-blood-based factors) substantially enhanced prediction capabilities compared to lifestyle factors alone, highlighting the synergistic effect of multi-domain feature integration.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp;Table 3\u0026nbsp;\u003c/strong\u003ehighlights top model performance per domain. The ensemble model with hybrid features achieved the highest overall performance using all features (AUROC 0.763, F1 0.728, accuracy 0.680), closely followed by XGBoost (AUROC 0.752). Among feature domains, metabolic and biochemical markers demonstrated the strongest predictive capability, with the ensemble model using original features achieving AUROC 0.759 and F1 0.729. Anthropometric and physical measurements showed moderate performance, with the ensemble hybrid approach reaching AUROC 0.679 and notably high recall (0.909). Lifestyle and socioeconomic factors alone had limited predictive power, with the best model (LGBM hybrid) achieving AUROC 0.595. When combining non-blood-based factors, the ensemble model with original features achieved AUROC 0.686 and F1 0.711, demonstrating the synergistic effect of integrating lifestyle and anthropometric data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2. Model Performance Comparison by Feature Domain and Dimensionality Reduction Strategy (AUROC)\u003c/strong\u003e\u003c/p\u003e\n\u003cdiv align=\"center\"\u003e\n \u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eFeature Domain\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003ePCA Strategy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eLR\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eRF\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eXGBoost\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eLGBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eANN\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eRNN\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eEnsemble\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\"\u003e\n \u003cp\u003e\u003cstrong\u003eAll Features\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eOriginal Features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.723\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.740\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.738\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.74\u003c/strong\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.637\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.691\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.75\u003c/strong\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid (Original+PC)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.725\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.735\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.752\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.750\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.601\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.693\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.763\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003ePC-Based\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.662\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.719\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.677\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.690\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.636\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.696\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\"\u003e\n \u003cp\u003e\u003cstrong\u003e(a) Metabolic and biochemical markers\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eOriginal Features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.707\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.731\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.746\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.73\u003c/strong\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.686\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.75\u003c/strong\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid (Original+PC)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.710\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.710\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.736\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.744\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.594\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.724\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.747\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003ePC-Based\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.632\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.637\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.641\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.657\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.656\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.639\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.671\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\"\u003e\n \u003cp\u003e\u003cstrong\u003e(b) Anthropometric and physical measurements\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eOriginal Features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.657\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.671\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.677\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.67\u003c/strong\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.611\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.588\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.67\u003c/strong\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid (Original+PC)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.659\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.662\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.652\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.686\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.637\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.613\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.679\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003ePC-Based\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.639\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.671\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.662\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.665\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.637\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.624\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.667\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\"\u003e\n \u003cp\u003e\u003cstrong\u003e(c) Lifestyle and socioeconomic factors\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eOriginal Features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.578\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.574\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.568\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.58\u003c/strong\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.531\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.555\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.58\u003c/strong\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid (Original+PC)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.580\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.572\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.586\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.595\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.533\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.515\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.583\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003ePC-Based\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.557\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.565\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.574\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.549\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.497\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.535\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.537\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\"\u003e\n \u003cp\u003e\u003cstrong\u003e(b+c) Non-blood-based factors\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eOriginal Features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.659\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.669\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.675\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.68\u003c/strong\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.592\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.570\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.68\u003c/strong\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid (Original+PC)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.660\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.653\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.671\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.678\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.618\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.560\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.670\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003ePC-Based\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.639\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.636\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.618\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.645\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.640\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.613\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e0.652\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u003csup\u003e*\u003c/sup\u003eBold indicates highest AUROC per domain; all values based on test set performance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3. Performance Metrics for Best Models in Each Domain\u003c/strong\u003e\u003c/p\u003e\n\u003cdiv align=\"center\"\u003e\n \u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"851\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eFeature Domain\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eBest Strategy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eAUROC\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eRecall\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eF1 Score\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eAll Features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eXGBoost\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.752\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.621\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.887\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.730\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.675\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eLGBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.750\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.640\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.843\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.728\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.687\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eEnsemble\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.763\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.631\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.860\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.728\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.680\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003e(a) Metabolic and biochemical markers\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eXGBoost\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eOriginal Features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.746\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.610\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.881\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.721\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.661\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eLGBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.744\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.622\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.885\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.731\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\"\u003e\n \u003cp\u003e0.676\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eEnsemble\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eOriginal Features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.759\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.634\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.859\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.729\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.683\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003e(b) Anthropometric and physical measurements\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eXGBoost\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eOriginal Features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.677\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.551\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.967\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.702\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.592\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eLGBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.686\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.564\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.943\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.705\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.609\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eEnsemble\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.679\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.574\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.909\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.704\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.620\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003e(c) Lifestyle and socioeconomic factors\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eXGBoost\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.586\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e1.000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.666\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.503\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eLGBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.595\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.504\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.985\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.667\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.511\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eEnsemble\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eHybrid\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.583\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.497\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e1.000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.664\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.497\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003e(b+c)\u0026nbsp;Non-blood-based factors\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eXGBoost\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eOriginal Features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.675\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.570\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.927\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.706\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.617\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eLGBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eOriginal Features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.686\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.557\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.942\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.700\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.599\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eEnsemble\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eOriginal Features\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.686\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.562\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.968\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.711\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e0.610\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u003csup\u003e*\u003c/sup\u003e Best strategy indicates the feature engineering approach that yielded the highest AUROC for each model within the respective domain.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3. Feature Importance Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4\u003c/strong\u003e presents the top 10 feature importance rankings for LGBM and XGBoost models across different feature domains. LGBM prioritized TG as the most important predictor, followed by ▲CRP and ALT, while XGBoost ranked ▲CRP first, followed by ▲ALT and TG. This algorithmic difference demonstrates distinct preferences, with XGBoost showing stronger emphasis on change variables compared to LGBM\u0026apos;s balance between static and dynamic measurements.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFigure 2\u003c/strong\u003e provides SHAP value distributions for the hybrid model using all features, offering deeper insights into how individual features contribute to T2D predictions. The SHAP analysis reveals distinct patterns in feature contributions, where ▲CRP (crp_diff) and TG show the widest impact ranges and highest feature values, confirming their dominant predictive influence observed in the rankings. For key metabolic markers, higher feature values (red points) generally shifted toward positive SHAP values, indicating that elevated levels of ▲CRP, TG, ALT, and ▲ALT increase T2D risk. The clear separation between high and low feature values in the SHAP distributions validates the ranking patterns observed in \u003cstrong\u003eTable 4\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003eThe convergence of evidence from both feature importance rankings and SHAP impact analyses consistently highlights several key findings. Change variables demonstrated superior predictive value for inflammatory and liver markers (▲CRP, ▲ALT), while static TG levels (pre-diabetic values) showed greater importance than TG changes (▲TG) in both ranking positions and SHAP impact magnitudes. The metabolic markers ▲CRP and TG not only achieved top rankings but also showed the strongest predictive impacts across both analytical approaches. This multi-faceted analysis confirms that temporal changes in key biomarkers provide the most robust signals for T2D risk assessment, with the SHAP distributions providing mechanistic validation of the feature importance hierarchies. risk assessment.\u003c/p\u003e\n\u003cp\u003eTo further validate the clinical interpretability of our models, we conducted detailed tree structure analysis of the XGBoost model. Examination of decision nodes revealed clinically meaningful split thresholds, such as BMI \u0026lt; 25.56 and waist circumference \u0026lt; 88.9 cm, which align closely with established clinical criteria for metabolic risk assessment (\u003cstrong\u003eSupplementary Methods S2.4\u003c/strong\u003e, \u003cstrong\u003eSupplementary Figure 4\u003c/strong\u003e). This analysis confirmed that the model was learning biologically plausible decision criteria rather than merely partitioning data mechanically, supporting the clinical applicability of our predictive framework. Complete tree structures and decision paths are provided in the supplementary materials to enhance reproducibility and transparency.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4. Top 10 Feature Importance Rankings by Domain in LGBM and XGBoost Models\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"1009\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eRank\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e2\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e3\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e4\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e5\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e6\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e7\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e8\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e9\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003e10\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003eAll Features\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eLGBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTG\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲CRP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eALT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eWBC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲TG\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eBio_\u003c/p\u003e\n \u003cp\u003ePCA_3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003ePhysical_\u003c/p\u003e\n \u003cp\u003ePCA_2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲WBC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲AST\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCRP\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eXGBoost\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲CRP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲ALT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTG\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eALT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003ePhysical_\u003c/p\u003e\n \u003cp\u003ePCA_2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲TG\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲WBC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲AST\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eBio_\u003c/p\u003e\n \u003cp\u003ePCA_2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Hematocrit\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003e(a) Metabolic and biochemical markers\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eLGBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTG\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲TG\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲CRP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲WBC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eAge\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCRP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eALT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Total_\u003c/p\u003e\n \u003cp\u003echolesterol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eBio_\u003c/p\u003e\n \u003cp\u003ePCA_10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eBio_\u003c/p\u003e\n \u003cp\u003ePCA_3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eXGBoost\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲CRP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTG\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲TG\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Hemoglobin\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲ALT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eALT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲AST\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲WBC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲RBC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Hematocrit\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003e(b) Anthropometric and physical measurements\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eLGBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Weight\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲BMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eWeight\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲WC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲DBP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003ePhysical_PCA_4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eBMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eWC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003ePhysical_\u003c/p\u003e\n \u003cp\u003ePCA_5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003ePhysical_\u003c/p\u003e\n \u003cp\u003ePCA_3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eXGBoost\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eBMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲BMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲SBP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Weight\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSBP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eResidential_area\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eWC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲WC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲DBP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSex\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003e(c) Lifestyle and socioeconomic factors\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eLGBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eAge\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Medical_cost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eLife_\u003c/p\u003e\n \u003cp\u003ePCA_8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Total_\u003c/p\u003e\n \u003cp\u003echolesterol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eLife_PCA_10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eLife_PCA_2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eResidential_area\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eLife_\u003c/p\u003e\n \u003cp\u003ePCA_4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eLife_\u003c/p\u003e\n \u003cp\u003ePCA_6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲pack_\u003c/p\u003e\n \u003cp\u003eyears\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eXGBoost\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSmoking\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Medical_cost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Smoking\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eResidential_area\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eAge\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Urinary_frequency\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲pack_\u003c/p\u003e\n \u003cp\u003eyears\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSex\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eLife_PCA_6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Total_\u003c/p\u003e\n \u003cp\u003echolesterol\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003e(b+c) Non-blood-based factors\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eLGBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eBMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Weight\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eWC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲BMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲WC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲SBP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eAge\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSBP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eWeight\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲DBP\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eXGBoost\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲BMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eBMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲SBP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Smoking\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Weight\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲ WC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eResidential_area\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eWC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSBP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e▲Urinary_frequency\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eAbbreviations: BMI, body mass index; TG, triglycerides; CRP, C-reactive protein; ALT, alanine aminotransferase; BUN, blood urea nitrogen; WBC, white blood cell count; SBP, systolic blood pressure; DBP, diastolic blood pressure; WC, waist circumference; Hb, hemoglobin; Tchl, total cholesterol; packyr_current, pack-years for current smokers. Features marked with ▲ represent change values (difference between follow-up Vₙ₋₁-Vₙ)\u003c/em\u003e, indicating increases from the previous visit.\u003c/p\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eThis study represents the first comprehensive investigation to develop machine learning models for predicting T2D onset by simultaneously incorporating both static and dynamic measurements of risk factors using population-based Korean cohort data. By employing an approach that excludes diagnostic glycemic markers while systematically considering longitudinal changes alongside static measurements in lifestyle factors and biomarkers, we present clinically relevant predictive models with significant translational potential.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWhile prospective cohort studies analyzing T2D risk among healthy adults are extremely limited, our findings align with recent machine learning approaches for T2D prediction, with our ensemble model achieving an AUROC of 0.763, comparable to the Fasa Adult Cohort Study (FACS) by Talebi Moghaddam et al. (AUC 89.61% in 7,408 Iranian adults over 5 years)\u003csup\u003e23\u003c/sup\u003e. Consistent with our findings, the FACS study demonstrated superior performance of ensemble methods, with Random Forest-based approaches achieving optimal results, reinforcing the robustness of tree-based algorithms for T2D prediction in prospective cohort studies \u003csup\u003e5\u003c/sup\u003e. A key distinction is our simultaneous incorporation of both static baseline measurements and longitudinal changes of risk factors, whereas the FACS study relied primarily on static measurements despite its prospective design. This temporal approach allows our model to capture dynamic metabolic risk evolution over time, potentially explaining disease progression rather than merely identifying high-risk individuals at baseline.\u003c/p\u003e\n\u003cp\u003eVarious machine learning algorithms have been utilized in previous T2D prediction studies \u003csup\u003e4-6,18\u003c/sup\u003e. Systematic reviews of diabetes prediction models revealed that the vast majority of studies (\u0026gt;90%) employed cross-sectional or retrospective designs, with prospective cohort studies being extremely rare \u003csup\u003e5\u003c/sup\u003e. Tree-based algorithms, particularly XGBoost, Random Forest, and ensemble methods, consistently achieved the highest performance across studies \u003csup\u003e4-6,18\u003c/sup\u003e. Cross-sectional studies with small, well-curated datasets (such as Pima Indians Diabetes Dataset \u003csup\u003e24\u003c/sup\u003e) achieved exceptional performance with AUC values of 0.85-0.95, largely attributed to the inclusion of diagnostic glycemic markers as predictive features, while studies using large-scale, real-world clinical datasets showed more modest performance with AUC ranges of 0.70-0.85 across diverse populations \u003csup\u003e4-6,18\u003c/sup\u003e. Notably, deep learning methods generally underperformed compared to tree-based algorithms despite their theoretical advantages for complex data patterns \u003csup\u003e25\u003c/sup\u003e. The significant performance gap between controlled datasets and real-world applications highlights the ongoing challenge of translating laboratory success to clinical practice.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eDomain-specific analysis revealed that metabolic and biochemical markers exhibited the most robust predictive power (Ensemble: AUROC 0.759), reflecting that T2D is fundamentally a metabolic disorder. Among these metabolic markers, change in C-reactive protein (▲CRP) emerged as the most important predictor in our feature importance analysis, highlighting the critical role of inflammatory processes in T2D pathogenesis. This finding aligns with previous longitudinal studies demonstrating that elevated CRP levels independently predict incident diabetes \u003csup\u003e26\u003c/sup\u003e, with a comprehensive meta-analysis of 22 cohorts involving 40,735 participants showing a 26% increased risk of T2D development in individuals with elevated CRP levels (RR 1.26; 95% CI 1.16-1.37) \u003csup\u003e27\u003c/sup\u003e. Notably, triglycerides (TG) ranked as the second most important predictor in LGBM models and third in XGBoost models, with static TG values demonstrating greater predictive importance than TG changes (▲TG), suggesting that baseline lipid levels may be more informative for diabetes risk assessment than dynamic lipid fluctuations. Both static BMI and BMI changes (▲BMI) showed significant predictive value, consistently ranking among the top predictors in non-blood-based factor analysis. Additionally, medical cost changes (▲Medical_cost) emerged as an important lifestyle predictor, potentially reflecting increased healthcare utilization patterns preceding T2D diagnosis and suggesting that healthcare usage patterns may serve as early warning signals for population screening strategies.\u003c/p\u003e\n\u003cp\u003eOur methodological approach introduces three distinct innovations that address key limitations in current T2D prediction research. First, we systematically evaluated multi-domain feature integration, demonstrating that combining non-blood-based factors (lifestyle + anthropometric; AUROC 0.686) achieved moderate predictive performance while representing modifiable risk factors that can guide practical interventions. Although this performance is substantially lower than biochemical markers (AUROC 0.759), these non-blood-based factors offer significant clinical value for early prevention strategies and risk screening in resource-limited settings where blood testing is not readily available. Second, we applied domain-stratified PCA within biologically meaningful feature groups (lifestyle, anthropometric, biochemical) rather than global dimensionality reduction across all variables \u003csup\u003e12,14\u003c/sup\u003e, which preserves biological interpretability while effectively addressing multicollinearity. Third, unlike conventional static risk assessment approaches, our temporal change-based features capture dynamic metabolic transitions preceding T2D onset, enabling detection of disease progression patterns rather than baseline risk stratification\u003c/p\u003e\n\u003cp\u003eThis study has several important limitations. First, our study was conducted in a single population cohort in South Korea, which may limit the generalizability of our findings to other populations with different genetic backgrounds, healthcare systems, and T2D risk profiles. Additionally, external validation was not feasible because other cohorts lacked the comprehensive diabetes diagnostic measurements (FBG, HbA1c, and 2-hour blood glucose levels after a 75g OGTT) that were consistently available in the Ansan-Ansung cohort, preventing us from applying our models to independent datasets for validation. Second, despite our rigorous exclusion criteria and comprehensive data cleaning procedures, residual confounding from unmeasured variables cannot be completely ruled out. Third, while our longitudinal design captures important temporal changes in biomarkers, the 2-year follow-up period may not fully capture the long-term dynamics of diabetes development, and longer follow-up studies would be valuable. Fourth, although we excluded participants with diagnostic glycemic indicators to avoid circular reasoning, this approach may have inadvertently excluded individuals in the very early stages of glucose dysregulation who might benefit most from early intervention. Finally, our feature selection process, while systematic, was limited to variables available in the cohort database. Notably, dietary data, which are crucial for diabetes prediction, could not be included as dietary surveys were conducted only twice during the study period (baseline and third follow-up), providing insufficient temporal resolution for our longitudinal modeling approach. This limitation potentially missed important nutritional and lifestyle factors that could enhance prediction accuracy.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn conclusion, dynamic changes in multiple risk factors over a 2-year period provided superior T2D predictive value compared to static measurements. The ensemble model achieved optimal performance (AUROC 0.763), closely followed by XGBoost (AUROC 0.752), with key predictors including inflammatory markers (▲CRP), lipid levels (TG), and anthropometric changes (▲BMI, ▲Weight). These findings offer significant potential for population-based screening, though external validation warrants further investigation.\u0026nbsp;\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eDeclaration of competing interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors have no competing interests to report.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data supporting the findings of this study are available from the Korean Genome and Epidemiology Study (KoGES) Ansan-Ansung cohort. Restrictions apply to the availability of these data, which were used under license for the current study. Data access requests can be submitted through the Korea Disease Control and Prevention Agency (KDCA) at https://coda.nih.go.kr/usab/koges/intro.do. Source code for data analysis and visualization is publicly available at https://github.com/tgkim-kr/MRSi-AI.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eConceptualization: HWW. Methodology: TK, HWW. Formal analysis: TK, JD, HWW. Writing - original draft: TK, HWW. Writing - review and editing: TK, JD, HWW.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-RS-2023-00246789)\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eCho, N. H. et al. IDF Diabetes Atlas: Global estimates of diabetes prevalence for 2017 and projections for 2045. \u003cem\u003eDiabetes Res. Clin. Pract.\u003c/em\u003e \u003cb\u003e138\u003c/b\u003e, 271\u0026ndash;281. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.diabres.2018.02.023\u003c/span\u003e\u003cspan address=\"10.1016/j.diabres.2018.02.023\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBanday, M. Z., Sameer, A. S. \u0026amp; Nissar, S. Pathophysiology of diabetes: An overview. \u003cem\u003eAvicenna J. Med.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e, 174\u0026ndash;188. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.4103/ajm.ajm_53_20\u003c/span\u003e\u003cspan address=\"10.4103/ajm.ajm_53_20\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSarkar, S. et al. The onset and the development of cardiometabolic aging: an insight into the underlying mechanisms. \u003cem\u003eFront. Pharmacol.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fphar.2024.1447890\u003c/span\u003e\u003cspan address=\"10.3389/fphar.2024.1447890\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNazirun, N. N. N. et al. Prediction Models for Type 2 Diabetes Progression: A Systematic Review. \u003cem\u003eIEEE Access.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e, 161595\u0026ndash;161619. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ACCESS.2024.3432118\u003c/span\u003e\u003cspan address=\"10.1109/ACCESS.2024.3432118\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFregoso-Aparicio, L., Noguez, J., Montesinos, L. \u0026amp; Garc\u0026iacute;a-Garc\u0026iacute;a, J. A. Machine learning and deep learning predictive models for type 2 diabetes: a systematic review. \u003cem\u003eDiabetol. Metab. Syndr.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, 148 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAsgari, S., Khalili, D., Hosseinpanah, F. \u0026amp; Hadaegh, F. Prediction Models for Type 2 Diabetes Risk in the General Population: A Systematic Review of Observational Studies. \u003cem\u003eInt. J. Endocrinol. Metab.\u003c/em\u003e \u003cb\u003e19\u003c/b\u003e, e109206. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.5812/ijem.109206\u003c/span\u003e\u003cspan address=\"10.5812/ijem.109206\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNaveed, I., Kaleem, M. F., Keshavjee, K. \u0026amp; Guergachi, A. Artificial intelligence with temporal features outperforms machine learning in predicting diabetes. \u003cem\u003ePLOS Digit. Health\u003c/em\u003e. \u003cb\u003e2\u003c/b\u003e, e0000354. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pdig.0000354\u003c/span\u003e\u003cspan address=\"10.1371/journal.pdig.0000354\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLugner, M., Rawshani, A., Helleryd, E. \u0026amp; Eliasson, B. Identifying top ten predictors of type 2 diabetes through machine learning analysis of UK Biobank data. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e14\u003c/b\u003e, 2102. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-024-52023-5\u003c/span\u003e\u003cspan address=\"10.1038/s41598-024-52023-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEjiyi, C. J. et al. A robust predictive diagnosis model for diabetes mellitus using Shapley-incorporated machine learning algorithms. \u003cem\u003eHealthc. Analytics\u003c/em\u003e. \u003cb\u003e3\u003c/b\u003e, 100166. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.health.2023.100166\u003c/span\u003e\u003cspan address=\"10.1016/j.health.2023.100166\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKim, Y. \u0026amp; Han, B. G. \u0026amp; group, t. K. Cohort Profile: The Korean Genome and Epidemiology Study (KoGES) Consortium. \u003cem\u003eInternational Journal of Epidemiology\u003c/em\u003e 46, e20-e20, (2016). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/ije/dyv316\u003c/span\u003e\u003cspan address=\"10.1093/ije/dyv316\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCommittee, A. D. \u0026amp; A. P., P. Committee:, A. D. A. P. P. 2. Classification and diagnosis of diabetes: standards of medical care in diabetes\u0026mdash;2022. \u003cem\u003eDiabetes care\u003c/em\u003e. \u003cb\u003e45\u003c/b\u003e, S17\u0026ndash;S38 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNilashi, M. et al. Knowledge Discovery and Diseases Prediction: A Comparative Study of Machine Learning Techniques. \u003cem\u003eJournal Soft Comput. \u0026amp; Decis. Support Systems\u003c/em\u003e \u003cb\u003e4\u003c/b\u003e (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGuyon, I. \u0026amp; Elisseeff, A. An introduction to variable and feature selection. \u003cem\u003eJ. Mach. Learn. Res.\u003c/em\u003e \u003cb\u003e3\u003c/b\u003e, 1157\u0026ndash;1182 (2003).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChangala, R. \u0026amp; Rao, D. R. Development of predictive model for medical domains to predict chronic diseases (diabetes) using machine learning algorithms and classification techniques. \u003cem\u003eARPN J. Eng. Appl. Sci.\u003c/em\u003e \u003cb\u003e14\u003c/b\u003e, 1202\u0026ndash;1212 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBreiman, L. \u0026amp; Random Forests \u003cem\u003eMach. Learn.\u003c/em\u003e \u003cb\u003e45\u003c/b\u003e, 5\u0026ndash;32, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1023/A:1010933404324\u003c/span\u003e\u003cspan address=\"10.1023/A:1010933404324\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2001).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen, T. \u0026amp; Guestrin, C. in \u003cem\u003eProceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining.\u003c/em\u003e 785\u0026ndash;794.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKe, G. et al. Lightgbm: A highly efficient gradient boosting decision tree. \u003cem\u003eAdvances neural Inform. Process. systems\u003c/em\u003e 30 (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAfsaneh, E., Sharifdini, A., Ghazzaghi, H. \u0026amp; Ghobadi, M. Z. Recent applications of machine learning and deep learning models in the prediction, diagnosis, and management of diabetes: a comprehensive review. \u003cem\u003eDiabetol. Metab. Syndr.\u003c/em\u003e \u003cb\u003e14\u003c/b\u003e, 196 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGardner, M. W. \u0026amp; Dorling, S. R. Artificial neural networks (the multilayer perceptron)\u0026mdash;a review of applications in the atmospheric sciences. \u003cem\u003eAtmos. Environ.\u003c/em\u003e \u003cb\u003e32\u003c/b\u003e, 2627\u0026ndash;2636 (1998).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMienye, I. D., Swart, T. G. \u0026amp; Obaido, G. Recurrent Neural Networks: A Comprehensive Review of Architectures, Variants, and Applications. \u003cem\u003eInformation\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e, 517 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOh, T. et al. Machine learning-based diagnosis and risk factor analysis of cardiocerebrovascular disease based on KNHANES. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e, 2250. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-022-06333-1\u003c/span\u003e\u003cspan address=\"10.1038/s41598-022-06333-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMujahid, M. et al. Data oversampling and imbalanced datasets: an investigation of performance for machine learning and feature engineering. \u003cem\u003eJ. Big Data\u003c/em\u003e. \u003cb\u003e11\u003c/b\u003e, 87. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s40537-024-00943-4\u003c/span\u003e\u003cspan address=\"10.1186/s40537-024-00943-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTalebi Moghaddam, M. et al. Predicting diabetes in adults: identifying important features in unbalanced data over a 5-year cohort study using machine learning algorithm. \u003cem\u003eBMC Med. Res. Methodol.\u003c/em\u003e \u003cb\u003e24\u003c/b\u003e, 220. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12874-024-02341-z\u003c/span\u003e\u003cspan address=\"10.1186/s12874-024-02341-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGupta, H., Varshney, H., Sharma, T. K., Pachauri, N. \u0026amp; Verma, O. P. Comparative performance analysis of quantum machine learning with deep learning for diabetes prediction. \u003cem\u003eComplex. Intell. Syst.\u003c/em\u003e \u003cb\u003e8\u003c/b\u003e, 3073\u0026ndash;3087. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s40747-021-00398-7\u003c/span\u003e\u003cspan address=\"10.1007/s40747-021-00398-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZou, Q. et al. Predicting diabetes mellitus with machine learning techniques. \u003cem\u003eFront. Genet.\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e, 515 (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePradhan, A. D., Manson, J. E., Rifai, N., Buring, J. E. \u0026amp; Ridker, P. M. C-reactive protein, interleukin 6, and risk of developing type 2 diabetes mellitus. \u003cem\u003eJama\u003c/em\u003e \u003cb\u003e286\u003c/b\u003e, 327\u0026ndash;334. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1001/jama.286.3.327\u003c/span\u003e\u003cspan address=\"10.1001/jama.286.3.327\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2001).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang, X. et al. Inflammatory markers and risk of type 2 diabetes: a systematic review and meta-analysis. \u003cem\u003eDiabetes Care\u003c/em\u003e. \u003cb\u003e36\u003c/b\u003e, 166\u0026ndash;175. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2337/dc12-0702\u003c/span\u003e\u003cspan address=\"10.2337/dc12-0702\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2013).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eManiruzzaman, M. et al. Accurate Diabetes Risk Stratification Using Machine Learning: Role of Missing Value and Outliers. \u003cem\u003eJ. Med. Syst.\u003c/em\u003e \u003cb\u003e42\u003c/b\u003e, 92. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s10916-018-0940-7\u003c/span\u003e\u003cspan address=\"10.1007/s10916-018-0940-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2018).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-7902454/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7902454/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eMost type 2 diabetes (T2D) prediction models rely on static baseline measurements and often include diagnostic glycemic markers, limiting their ability to capture temporal risk evolution and creating circular reasoning. This study developed a machine learning framework that systematically integrates baseline measurements with longitudinal interval changes to predict incident T2D. Using the Ansan-Ansung cohort of the Korean Genome and Epidemiology Study (KoGES; 2001\u0026ndash;2018), we included 7,510 initially diabetes-free participants in this prospective analysis. The framework jointly modeled static variables and 2-year interval changes in lifestyle, anthropometric, and biochemical markers using XGBoost, Random Forest, LGBM, logistic regression, neural networks, and ensemble methods. Principal component analysis addressed multicollinearity. Diagnostic glycemic markers (fasting glucose, HbA1c) were excluded to ensure genuine risk prediction. The ensemble model achieved AUROC 0.763, with XGBoost (0.752) and LGBM (0.750) showing comparable performance. SHAP analysis identified changes in C-reactive protein (▲CRP) and body mass index (▲BMI), together with baseline triglycerides, as the most influential predictors. Examination of decision tree structures revealed clinically meaningful and biologically plausible thresholds (e.g., BMI\u0026thinsp;\u0026lt;\u0026thinsp;25.6 kg/m\u0026sup2;). The resulting ensemble model was implemented through the Multi-Domain Simulation Interface (MDSi) framework, enabling population-level inference across lifestyle, anthropometric, and metabolic domains. Overall, change variables contributed more strongly than static measures, suggesting that accelerated physiological shifts precede the onset of T2D. By capturing dynamic metabolic trajectories rather than static risk profiles, this framework differentiates true risk prediction from early disease detection, enabling clinically interpretable prediction with substantial potential for preventive interventions before diagnostic thresholds are reached.\u003c/p\u003e","manuscriptTitle":"Predicting Type 2 Diabetes Using Baseline and Longitudinal Changes in Lifestyle and Clinical Markers: A Machine Learning Approach","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-05 01:50:46","doi":"10.21203/rs.3.rs-7902454/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2025-12-22T23:17:33+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"199638595560365486793799538930464318026","date":"2025-12-02T14:18:59+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"140197207819964418712034905912733818412","date":"2025-12-01T12:19:26+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-12-01T11:36:25+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-10-30T06:02:12+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-10-23T10:01:43+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-10-23T09:59:49+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-10-20T05:47:19+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"4d93bdfb-0488-4983-9d90-78753faa4ae4","owner":[],"postedDate":"December 5th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":58990460,"name":"Health sciences/Biomarkers"},{"id":58990461,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":58990462,"name":"Health sciences/Diseases"},{"id":58990463,"name":"Health sciences/Endocrinology"},{"id":58990464,"name":"Health sciences/Health care"},{"id":58990465,"name":"Health sciences/Medical research"},{"id":58990466,"name":"Health sciences/Risk factors"}],"tags":[],"updatedAt":"2025-12-05T01:50:46+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-05 01:50:46","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7902454","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7902454","identity":"rs-7902454","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}