This study enrolled 4,056 T2DM individuals who received health care through routine physical examinations, outpatient visits, and inpatient admissions at the First Affiliated Hospital of Ningbo University, Ningbo, China, from November 2019 to November 2024. Demographic data, relevant complications, biochemical parameters, etc., were obtained through questionnaire surveys and laboratory examinations. Ultimately, 1,553 participants were included in the study. The specific exclusion criteria were as follows: (1) patients without liver ultrasound results; (2) patients with body mass index (BMI) ≤ 18 or BMI ≥ 24; (3) patients with heavy alcohol consumption (exceeding 140 grams per week for men and 70 grams per week for women); (4) diagnosis of liver disease, such as viral hepatitis, and autoimmune hepatitis. We used the multiple imputation method (linear regression, polynomial logistic regression, and five iterations to create an interpolation model) to process missing values, in order to reduce bias toward missing data ( Supplementary Figure 1 ) (25). The study’s flow diagram is depicted in Figure 1 .

The diagnostic criteria for DM were fasting blood glucose (FBG) levels of ≥ 7.0 mmol/L, 2-hour blood glucose levels of ≥ 11.1 mmol/L, or a glycated hemoglobin level of ≥ 6.5% (26).

Clinical baseline data encompassed the Participants’ general characteristics (gender, age, BMI, systolic blood pressure [SBP], diastolic blood pressure [DBP], and heart rate [HR]), lifestyle habits(history of smoking), blood cell counts (white blood cell count [WBC], red blood cell count [RBC], mean red blood cell volume [MCV], lymphocyte count [LYMPH], monocyte count[MONO], neutrophil count [NEUT], eosinophil count [EOS], basophil count [BASO], hemoglobin [HB], platelet count [PLT], mean platelet volume [MPV], and platelet distribution width [PDW]), biochemical indicators (alanine aminotransferase [ALT], aspartate aminotransferase [AST], gamma-glutamyl transpeptidase [GGT], total bilirubin [TBIL], direct bilirubin [DBIL], indirect bilirubin [IBIL], albumin [ALB], globulin [GLO], total protein [TP], creatinine [CREA], uric acid [UA], total cholesterol [TC], triglycerides [TG], high-density lipoprotein cholesterol [HDL-C], low-density lipoprotein cholesterol [LDL-C], and glucose [GLU]), and other laboratory value(glycosylated hemoglobin [HBA1C]).

The preliminary analysis of the dataset involves the application of descriptive statistics. The Kolmogorov-Smirnov (K-S) test was used to test the normality of the continuous variables. Continuous variables with a normal distribution were expressed as means (standard deviations, SD), continuous variables with a skewed distribution were expressed as medians (interquartile ranges, IQR), and categorical variables were expressed as percentages (percentage, %). Independent-samples T test (continuous variables with a normal distribution), Mann-Whitney U test (continuous variables with a skewed distribution), and chi-square test (categorical variables) were used to evaluate the differences between groups, and the standardized mean difference (SMD) was used to evaluate balance between groups (27).

We used multivariate logistic regression to identify independent risk factors for lean NAFLD. The Boruta algorithm and the Least Absolute Shrinkage and Selection Operator (LASSO) were utilized to screen for characteristic variables. To construct a prediction model, the entire dataset was partitioned into a training set and an internal validation set at a ratio of 6:4. Subsequently, six ML algorithms including linear discriminant analysis (LDA), logistic regression (LR), Naive Bayes (NB), random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGboost), were employed to train the model. During the model training process, a 10-fold cross-validation method was utilized to optimize the model parameters and prevent the occurrence of overfitting. In addition, we employed the AUC to evaluate the predictive ability of the model. Calibration curves and the Brier score were utilized to assess the calibration ability, while decision curve analysis (DCA) was applied to evaluate the clinical applicability. Additionally, the Shapley Additive exPlanations (SHAP) was used to interpret the best predictive model.

Statistical analyses were conducted using the R language (version 4.2.3, http://www.R-project.org/) and Python (version 3.9.0, https://www.python.org/). All data were analyzed using two-sided tests, and statistical significance was defined as P < 0.05.

A total of 1,553 participants were recruited in this study, including 1,237 T2DM patients without lean NAFLD and T2DM 316 patients with complicated by lean NAFLD. The median age of the patients was 59.00 years (IQR:50.00–67.00), among whom 713 cases (45.9%) were male and 840 cases (54.1%) were female. When comparing the baseline characteristics between the two groups of patients, statistically significant differences were observed in terms of gender, BMI, DBP, routine blood tests, liver and kidney functions, blood lipid levels, and blood glucose levels (p < 0.05). As shown in Table 1 .

Figure 2 presented the prevalence rate of lean NAFLD in the study population. The overall prevalence rate of lean NAFLD is 20.3%, with the prevalence rate among females (20.5%) being slightly higher than that among males (20.2%) ( Figure 2A ). Age pattern analysis showed that the prevalence of lean NAFLD decreased with increasing age ( Figure 2B ).

Twenty potential risk factors associated with lean NAFLD were screened through univariate analysis (p < 0.05 and SMD > 0.1) ( Table 1 ). To ensure the accuracy and reliability of the research results, the variance inflation factor (VIF) of each variable was calculated. Variables with a VIF value exceeding 10 typically exhibit severe multicollinearity ( Supplementary Table 1 ). Subsequently, we employed stepwise backward logistic regression analysis with the Akaike information criterion to filter and remove multicollinear variables. Ultimately, twelve variables were included for the multivariate logistic regression analysis, and eight variables such as Age, BMI, ALT, GGT, IBIL, ALB, UA, and TG were identified as independent risk factors for lean NAFLD (p < 0.05). The results were presented in Figure 3 .

The study population was divided into a training set and an internal validation set at a ratio of 6:4. Statistical analysis revealed no significant differences between the two groups ( Table 2 ).

LASSO regression is a data reduction method that reduces the complexity of the model, prevents overfitting, and selects important feature variables by formulating an optimized objective function with a penalty term (28). In this study, 22 characteristic factors were identified by using LASSO regression ( Figure 4A ).

The Boruta algorithm is a feature selection method based on random forests, aiming to identify truly significant features from a given feature set and distinguish irrelevant features (29). Thirteen key factors were identified through the Boruta algorithm ( Figure 4B ).

In the training set, through a comparative analysis of the screening results of the LASSO regression and the Boruta algorithm, we identified the common subset of feature variables selected by the two methods ( Figure 4C ). These selected variables were used as predictors to construct a risk prediction model for lean NAFLD, including BMI, LYMPH, HB, ALT, AST, GGT, IBIL, ALB, UA, and TG. In addition, the optimal model was determined among the risk prediction models constructed by six machine learning algorithms, namely LDA, LR, NB, RF, SVM, and XGboost.

Within the training set, the RF model exhibited outstanding predictive performance [AUC: 0.739 (95%CI: 0.676–0.802)]. In contrast, the AUC values of the remaining five models were as follows: 0.723 (95%CI: 0.682–0.764) for LDA, 0.723 (95%CI: 0.691–0.755) for LR, 0.694 (95%CI: 0.647–0.741) for NB, 0.635 (95%CI: 0.579–0.727) for SVM, and 0.733 (95%CI:0.69–0.776) for XGboost ( Figure 5A ). In the internal validation set, the RF model also demonstrated robust clinical predictive value [AUC: 0.789 (95%CI: 0.722–0.856)] ( Figure 5B ). Furthermore, we conducted a comprehensive comparative analysis of additional clinical performance metrics, including accuracy, sensitivity, specificity, precision, recall, and F1 score, across various predictive models in both the training set and internal validation set ( Tables 3 , 4 ). The table revealed that the RF model demonstrates superior performance across all evaluated parameters. Concurrently, cut-off values were respectively established for the predictive probabilities of six ML models in both the training set and internal validation set ( Figure 6 ). Patients were classified as positive if the predicted probability exceeded the cut-off value; otherwise, they were categorized as negative. Consequently, confusion matrixes of the predicted probabilities and the actual values were plotted, as illustrated in the Figures 7 , 8 .

In this study, we evaluated the consistency between the model’s predicted probabilities and the actual occurrence probabilities through the analysis of calibration curves for the training set and the internal validation set. As shown in Figure 9 , in both the training set and the internal validation set, except for the NB and XGboost models, the predicted values of the remaining models were in good agreement with the theoretical values, indicating favorable clinical calibration. Among them, the Brier score of RF model was the smallest, which reflected the high reliability of the model’s prediction.

The DCA curve assessed the clinical decision-making value of the model under different thresholds through the net benefit. In the training set and internal validation set, except for the NB and XGboost models, the remaining models (especially the RF model) exhibited favorable clinical decision-making value. The results were presented in Figure 10 . We further calculated the risk threshold probabilities of the RF model in the training set. The results indicated that when the threshold probability ranged between 1% and 55%, the net benefit provided by the model was significantly higher than that of the baseline strategy. In the internal validation set, the model also exhibited favorable net benefits, particularly demonstrating clinical advantages within the threshold probability range of 1% to 50%.

ML models achieve favorable performance by capturing data patterns through intricate mathematical structures. However, their complexity makes it difficult to interpret their internal decision - making processes, and they are commonly regarded as “black-box” models. SHAP is a tool designed for interpreting machine learning models. It explains the prediction results of a model by assigning a “contribution value” (Shapley value) to each feature, thereby rendering the decision-making process of black-box models transparent and controllable (30). Based on the above results, the RF model emerged as the optimal predictive model, demonstrating stable performance in both the training set and the internal validation set. Therefore, we further employed interpretable tools to analyze the contributions of characteristic predictors to the RF model. Figure 11A illustrated the contribution of characteristic predictors to the prediction model in the RF model, where TG, ALT, GGT, BMI, and UA were the top five variables in terms of importance. From Figure 11B , we could observe that the larger the TG, ALT, GGT, BMI, and UA, the greater the SHAP value and the higher the risk of disease.