The study population was selected from individuals who underwent health checkups at the Health Management Center of the First Affiliated Hospital of Soochow University. Data on health check-ups from January 2018 to December 2022 were collected for this study. To ensure the accuracy and reasonableness of the data, the research paid special attention to the last five years of data from the Health Management Center. According to the design of a retrospective cohort study, this study considered the 2018 health check-up data as the baseline of the cohort and the study ended in 2022. The presence of impaired fasting glucose (IFG) in this study population was the primary outcome of interest.

The inclusion criteria were as follows: (1) aged 45 years or older;(2) not previously diagnosed with IFG or T2DM at baseline;(3) possessed complete physical examination data from 2018 to 2022 without significant gaps in critical information. Exclusion criteria included: (1) age less than 45 years; (2) prior or recent diagnosis of IFG or T2DM at baseline; (3) use of medications that could influence plasma glucose levels;(4)missing or incomplete data on key clinical parameters such as fasting plasma glucose (FPG).

Demographic and sociological information, including age, gender, and medical history, was collected from all participants in this study. In addition, the results of each participant’s physical examination were recorded and laboratory measurements were taken. This involved measuring height, weight, systolic and diastolic blood pressure (SBP and DBP) and calculating the body mass index (BMI) based on the participant’s height and weight. All participants underwent fasting for a minimum of 8 hours before morning examinations, during which 3-5ml of venous blood was drawn from the elbow. Fasting plasma glucose (FPG), blood urea nitrogen (BUN), serum uric acid (SUA), serum creatinine (Scr), alanine aminotransferase (ALT), aspartic acid aminotransferase (AST), glutamyl transpeptidase (GGT), alkaline phosphatase (ALP), total cholesterol (TC), triglycerides (TG), high-density lipoprotein cholesterol (HDL), low-density lipoprotein cholesterol (LDL), apolipoprotein A1 (ApoA1), and apolipoprotein B (ApoB) were assessed using the Hitachi 7600 Automatic Biochemistry Analyzer. Furthermore, abdominal ultrasound examinations were performed by trained sonographers, and all participants underwent abdominal ultrasound scans. To accommodate machine learning algorithms that require numeric feature attributes, non-numeric attributes like gender were converted into numeric values. The percentage of missing variables included in this study was less than 30%. The MICE package was used for missing value analysis and multiple interpolation. Supplementary Figure 1 illustrates the results of data interpolation. The flowchart of this study is shown in Figure 1 .

In accordance with the World Health Organization (WHO) diagnostic criteria for diabetes mellitus established in 1999 (14), the FPG levels of the participants were classified in this study. A normal blood glucose state was characterized by a fasting glucose level below 6.10 mmol/L and a 2-hour glucose level from the oral glucose tolerance test (OGTT) under 7.80 mmol/L. IFG was defined as a fasting glucose level ranging from 6.10 to 7.00 mmol/L or a 2-hour OGTT glucose level below 7.80 mmol/L, not meeting the diagnostic criteria for diabetes. For the diagnosis of fatty liver, we adhered to the ultrasound criteria outlined in the National Workshop on Fatty Liver and Alcoholic Liver Disease, Chinese Society of Hepatology, Chinese Medical Association (15). In this study, the diagnosis of fatty liver was confirmed using ultrasound diagnostic criteria proposed at the National Symposium on Fatty Liver and Alcoholic Liver Disease of the Chinese Medical Association Hepatology Branch (15). Specifically, fatty liver diagnosis relied on one of the following criteria: (1) noticeable enhancement of near-field liver echoes surpassing that of the kidney; (2) indistinct intrahepatic duct structure; (3) gradual attenuation of liver echoes in the far field. Gallstones were identified through ultrasound imaging, typically presenting as one or more intense echoes within the gallbladder, extrahepatic bile duct, or intrahepatic bile duct, accompanied by movable acoustic shadows (16). A history of cholecystectomy denoted prior gallbladder removal, where the gallbladder was no longer visible on ultrasound (17). The term “gallstones” encompassed both the presence of gallstones and prior cholecystectomy (18).

The data collected from the medical checkup center of the First Affiliated Hospital of Soochow University were randomly divided into a training set and a validation set at a ratio of 8:2. Feature selection, parameter tuning, and model building were performed in the training set, while the validation set was used to evaluate the predictive performance of the models. Variable selection was the main step before the modelling. Boruta algorithm belongs to one of the random forest algorithms, its main purpose is to screen and sort important characteristic variables related to dependent variables. In each iteration, a comparison of the importance of the original and shadow variables is performed. If the importance of the original variable is significantly higher than the importance of the shadow variable, the original variable is considered important; if the importance of the original variable is significantly lower than the importance of the shadow variable, the original variable is considered unimportant (19). Boruta algorithm reaches a specified limit of random forest operation after 99 iterations. All variables were selected by Boruta algorithm and then multivariate logistic regression analysis was performed. The filtered variables were incorporated into the ML model. ML prediction models were developed, comprising XGBoost, SVM, Naive Bayes, DT and LR models. To determine the optimal parameters, a 5-fold cross-validation grid search was executed on the training set, while the LR model was implemented with default parameters.

Based on the prediction results of the model, the area under the receiver operating characteristic curve (AUC) of the training and validation sets can be calculated to assess the discriminative ability of the model. The confusion matrix, consisting of true positives (TP), true negatives (TF), false positives (FP), and false negatives (FN), was established to calculate Sensitivity, Specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy (ACC), and F1 scores for evaluating discrimination performance of models. Formulas were as follows: Sensitivity=Recall= TP/(TP+FN); Specificity= TN/(TN+FP); ACC = (TP + TN)/(TP + FP + FN + TN); PPV = Precision = TP/(TP + NP); NPV = TN/(TN + FN); F1 score= 2*Recall*Precision/(Recall +Precision). Calibration curves plotted after sampling with repetition 500 times using the Bootstrap method reflect the fitting of the models. The decision curves (DCA) were used to assess the clinical utility of the predictive model. To gain insight into the factors that contribute to the development of IFG in middle-aged and elderly people, SHapley Additive Explanations (SHAP) plots were further developed in this study. These plots visually demonstrated the variable contributions to the outcome, with local SHAP plots providing a detailed look at variable contributions for specific instances. Feature importance was ranked according to the Shapley value. In addition, force plots within the SHAP model were used to individualize predictions for two randomly selected samples from the validation set.

Statistical analysis was performed using SPSS (version 27.0) to describe the general data of all individuals. The Shapiro–Wilk test was employed to ascertain the normal distribution of variables. Continuous variables displaying a normal distribution were presented as mean ± standard deviation (SD), while those with skewed distribution were described as median (interquartile ranges). Categorical variables were represented as frequencies. To compare variables between groups, the Pearson Chi-square test was used for categorical variables, while the Student’s t-test or the nonparametric Mann-Whitney U test was applied for continuous variables. All variables, determined as significant and tentative through Boruta algorithm screening, were incorporated into the multivariate logistic regression analysis. The logistic regression analysis was conducted using Enter approach, with α in = 0.05, α out = 0.1, and an inspection level of α = 0.05. SPSS (version 27.0) was also used to draw box plots to depict the independent risk factors associated with the onset of IFG in the middle-aged and elderly people identified in this study. The data interpolation, feature selection, model construction, evaluation, and visualization were executed using R software (version 4.2.3). The main packages involved are “mice, Boruta, caret, xgboost, shap for xgboost, shapviz, Resource Selection, rms”. A two-sided p < 0.05 was considered statistically significant.

A total of 1,855 individuals were included in the cohort, with 734 cases (39.0%) of impaired fasting glucose (IFG) observed within the entire cohort. They were randomly split into training (n = 1,438) and validation (n = 417) sets in an 8:2 ratio. In the training dataset, 73.0% (1,050/1,438) were male, and 27.0% (388/1,438) were female. The median age was 56 years (IQR = 50–67 years) for the IFG group and 55 years (IQR = 49–65 years) for the non-IFG group. In the test dataset, IFG onset was more common among male patients, with median ages ranging from 60 to 71 years. Detailed demographic and clinical characteristics are presented in Table 1 .

Following 99 iterations, the Boruta algorithm’s feature variable screening results are illustrated in Figure 2 . The following variables were considered important for their association with IFG: SBP, Fatty liver, BMI, WC, TG, Scr, Age, DBP, ApoB, TC, SUA and GGT. To further clarify the risk or protective factors related to IFG, multivariate regression analysis was conducted between the IFG and non-IFG groups in the training cohort, revealing significant differences in the following variables: SBP, Fatty liver, WC and Scr. This study confirmed that Scr was a protective factor associated with IFG, while SBP, Fatty liver and WC were all risk factors for IFG. The results showed that for every 1 mmHg increase in SBP, the risk of IFG in middle-aged and elderly people increased by 3.0% (OR =1.030, 95% CI: 1.020-1.040). The risk of IFG in middle-aged and elderly people increased by 3.3% for every 1 cm increase in waist circumference (OR = 1.034, 95% CI: 1.009-1.059). The risk of IFG was 50.5% higher in middle-aged and elderly people with fatty liver compared to those without fatty liver disease (OR = 1.657, 95% CI: 1.274-2.156). The risk of IFG in middle-aged and elderly people was elevated by 1.6% for every 1μmol/L decrease in Scr levels (OR = 0.984, 95% CI: 0.975-0.993). Further details are provided in Table 2 . In the study, box plots describing the distribution of these three continuous variables including SBP, WC and Scr in the IFG and non-IFG groups were further plotted based on the training set. As shown in Figure 3 , middle-aged and elderly people who developed IFG over a 5-year period had higher SBP, larger WC and lower Scr than middle-aged and elderly people without IFG.

Four features selected by the Boruta algorithm and multivariate regression analysis were considered as input variables, with the development of IFG as the outcome. Six different algorithms, including XGBoost, SVM, Naïve Bayes, DT and LR, were applied to construct prediction models in this study. The AUC, Sensitivity, Specificity, ACC, PPV, NPV, and F1 scores of the model in the validation set were calculated from the confusion matrix results. The details are showed in Table 3 . Among these models, the XGBoost model yielded the highest AUC, indicating superior performance. With the aid of grid search, the optimal structure of the XGBoost model was determined as follows: booster=‘gbtree’, objective=‘binary:logistic’,gamma=0.5, eta=0.06,max_depth=7, min_child_weight=5, subsample=0.65, colsample_bytree=0.72. Table 3 displays the superior AUC value achieved by the XGBoost algorithm compared to SVM, Naive Bayes, DT and LR algorithms (0.7391, 0.7328, 0.7288, 0.6480 and 0.6795) respectively. Typically, a model with an AUC greater than 0.7 is considered to have good predictive performance. The calibration curves for the training and validation sets are plotted in this study and are shown in Figure 4 . The results demonstrate the internal validation of the XGBoost model using the bootstrap method with 500 repetitions of sampling. The mean absolute errors for the training and validation sets were 0.010 and 0.025, respectively, indicating that the predicted probabilities of the XGBoost model closely aligned with the actual observations. The results of the Hosmer-Lemeshow goodness of fit test(H-L) showed that the model was well fitted (p>0.05). The clinical effect of the XGBoost prediction model was evaluated in this study using DCA curves, as shown in Figure 5 , showing that individuals with a higher risk of developing IFG in middle-aged and elderly people as assessed using this XGBoost model may have a higher net benefit value if they were intervened. In addition, the SHAP framework has provided an intuitive interpretation of the XGBoost model, as shown in Figure 6 .

Figure 7 illustrates the SHAP analysis, showcasing the role of important variables on individual predictions in 2 randomly selected samples from the validation set. The local SHAP plots illustrated the contributions of variables to the outcomes for each sample. In Figure 7 , the purple portion of the local SHAP force plot represents support for a positive prediction, while the yellow portion indicates support for a negative prediction. The length of the feature lines corresponds to the size of their contribution. For example, case 1 exhibited a high predicted probability of 0.82 for progressing to IFG, as predicted by the XGBoost model. SBP was the most significant feature contributing to the prediction, followed by Scr, Fatty liver and WC. This was enough to confirm the usefulness of the XGBoost model and contribute to increasing doctors’ trust in the predictive model to help them make the right auxiliary decisions.