Integrating body composition analysis and machine learning for non-invasive identification of metabolic dysfunction-associated fatty liver disease: a large-scale health examination-based study

Abstract

Metabolic dysfunction-associated fatty liver disease (MAFLD) is a highly prevalent liver condition closely linked to obesity, insulin resistance, and metabolic syndrome. Early identification of MAFLD remains challenging in routine health examination settings remain challenging, especially in routine health examination settings where conventional indicators often fail to capture deeper metabolic disturbances. This study aimed to evaluate the predictive value of body composition parameters and develop and validate a non-invasive, machine learning-based classification model for MAFLD. A retrospective study was conducted using data from 23,348 adults who underwent health check-ups between 2017 and 2021 at a tertiary hospital in China. Body composition was assessed via bioelectrical impedance analysis, and MAFLD was diagnosed based on hepatic steatosis plus metabolic risk criteria. A total of 13 features, including body composition indicators and basic demographics, were initially considered. Feature selection was guided by multicollinearity diagnostics and model-based importance analysis. Eight machine learning models were constructed and evaluated using tenfold cross-validation. An independent external validation cohort of 3,357 participants from 2022 to 2023 was used to assess generalizability. Performance was evaluated using area under the receiver operating characteristic curve, accuracy, recall, F1 score, and calibration metrics. Among all models, tree-based algorithms including extreme gradient boosting, gradient boosting decision tree, and LightGBM achieved the highest discriminative performance, with internal validation area under the curve values exceeding 0.96 and external validation area under the curve values above 0.95. Visceral fat rating consistently emerged as the most important predictor, followed by waist circumference and body mass index. Logistic regression confirmed their independent associations with MAFLD after adjustment for key confounders. Stratified analyses revealed variable patterns across sex, age, and body mass index groups, with visceral fat remaining a robust predictor in all subgroups. Body composition analysis, particularly visceral fat estimation, demonstrates strong diagnostic discrimination for MAFLD using non-invasive measurements. Integrating these parameters with machine learning enables accurate identification, supporting scalable screening and aiding diagnostic assessment in routine health examination, clinical, and public health settings.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-37852-w.

Introduction

Research insights

What is currently known about this topic?

Visceral fat is a key driver of metabolic disorders such as metabolic dysfunction-associated fatty liver disease (MAFLD).

Traditional markers like body mass index (BMI) and waist circumference (WC) are insufficient for accurate prediction.

Bioelectrical impedance analysis (BIA) is a non-invasive method for assessing body composition and metabolic risk.

What is the key research question?

Can BIA-derived parameters predict MAFLD risk more accurately using machine learning models?

What is new?

Machine learning models using BIA parameters achieved excellent predictive performance.

Visceral fat and extracellular water percentage (ECW%)were identified as key independent predictors of MAFLD.

Model generalizability was validated in a large external health examination cohort.

How might this study influence clinical practice?

Supports non-invasive, scalable screening strategies for early MAFLD detection.

Metabolic dysfunction-associated fatty liver disease (MAFLD), formerly known as nonalcoholic fatty liver disease (NAFLD), is a globally prevalent chronic liver disease recently redefined to emphasize metabolic dysregulation as its core feature^1,2. It is estimated that MAFLD affects approximately one-quarter of the adult population worldwide, with prevalence varying across countries and regions³. In the Chinese population undergoing routine health check-ups, the overall prevalence of MAFLD reaches 37%, with a significantly higher rate in men (46%) compared to women (24%), and a clear upward trend with advancing age⁴. In urban community settings, the prevalence of MAFLD diagnosed via abdominal ultrasonography has been reported to be as high as 26%⁵. MAFLD is closely associated with obesity, type 2 diabetes mellitus (T2DM), and dyslipidemia, and markedly increases the risk of cardiovascular disease, chronic kidney disease, and multiple systemic metabolic complications, thereby posing a substantial public health and economic burden^6,7.

The redefinition of NAFLD as MAFLD underscores the pivotal role of metabolic health in the onset and progression of hepatic steatosis. Despite growing attention to MAFLD, its early identification remains a significant challenge, particularly among individuals undergoing routine health examinations, where precise risk stratification of high-risk individuals is often difficult^8,9. Existing studies primarily focus on conventional biochemical markers such as liver enzymes, lipid profiles, and glucose levels; however, these parameters fail to capture the underlying metabolic disturbances, including ectopic fat accumulation, loss of skeletal muscle mass, and imbalances in body fluid distribution^10,11. Although imaging techniques and liver biopsy represent diagnostic gold standards, they are limited by their invasiveness, high cost, and limited accessibility in primary care settings. Therefore, there is an urgent need for a non-invasive, easily obtainable, and metabolically informative alternative tool to facilitate early detection and identification of MAFLD.

Against this background, bioelectrical impedance analysis (BIA), a rapid and non-invasive technique for assessing body composition, has garnered increasing attention^12,13. BIA provides a wide range of parameters, including fat mass, muscle mass, fluid distribution, and visceral fat rating, all of which are closely linked to metabolic health^14,15. Although the utility of BIA has been partially validated in the context of obesity and metabolic syndrome, large-scale population-based studies systematically evaluating its predictive performance for MAFLD risk remain limited^16,17. In particular, it is unclear whether BIA-derived parameters retain independent predictive value for MAFLD after adjusting for key confounders¹⁸. Moreover, it is necessary to investigate whether the predictive capacity of BIA indices varies across different subgroups stratified by sex, age, or body mass index (BMI).

Given the increasing global burden of MAFLD and its profound impact on health, early diagnosis and individualized risk assessment remain major clinical challenges. Therefore, the present study aims to systematically evaluate the associations between body composition parameters and MAFLD status, and to develop an efficient machine learning-based diagnostic identification model to enhance early screening capabilities. This research is expected to improve non-invasive screening and personalized risk stratification for MAFLD by providing a simple, accurate, and clinically applicable tool. Ultimately, it may offer theoretical support for individualized treatment and early intervention strategies, thereby helping to reduce the public health burden of the disease. The study design is illustrated in Fig. 1.

Fig. 1.

Flowchart of Data Processing, Model Construction, and External Verification for MAFLD Identification Based on Machine Learning.

The training set (n = 23,348) from 2017.01 to 2021.12 was used to construct prediction models using various algorithms, followed by tenfold cross-validation. The external verification set (n = 3,357) from 2022.01 to 2023.12 was used to assess model generalizability. MAFLD diagnosis was based on international consensus criteria. Evaluation metrics included accuracy, precision, recall, F1 score, and ROC AUC.

Methods

Study design and data sources

This retrospective cross-sectional study was conducted using data from individuals who underwent routine health check-ups at the Health Management Center of the Third Xiangya Hospital, Central South University. All data were derived from previous examinations, with no additional interventions or sampling performed. All analyses were conducted in accordance with the ethical principles of the Declaration of Helsinki. Participant data were anonymized to protect privacy and ensure data security. As the study involved retrospective data analysis, the requirement for written informed consent was waived by the institutional ethics committee.

A total of 28,663 individuals who completed both comprehensive body composition analysis and abdominal ultrasound between January 2017 and December 2021 were initially assessed. After applying exclusion criteria (shown in Fig. 2), a final sample of 23,348 individuals was included in the study. Additionally, health check-up data from 4,379 individuals who underwent InBody body composition analysis between January 2022 and December 2023 at the same center were collected. After applying the inclusion and exclusion criteria (see Supplementary Fig. S1), 3,357 participants were included as an external validation cohort. This external cohort, which also comprised individuals who had undergone InBody assessments, was used to evaluate the generalizability and stability of the machine learning models developed in this study across different populations.

Fig. 2.

Flowchart of Participant Selection for the Model Development Cohort (2017–2021). A total of 28,663 individuals who completed body composition testing and abdominal ultrasound were initially enrolled. After applying age restrictions and additional exclusion criteria (e.g., alcohol abuse, liver or kidney disease, malignancy, pregnancy), 23,348 subjects were included in the final analysis, comprising 9,479 MAFLD cases and 13,869 non-MAFLD controls.

Variable selection and measurements

All measurement data were collected by trained researchers at the Third Xiangya Hospital. Baseline demographic characteristics, lifestyle behaviors, medical history, family history, and medication use were obtained through face-to-face interviews using standardized questionnaires. Physical examinations were conducted to measure height, weight, waist circumference, and blood pressure. BMI was calculated as weight in kilograms divided by height in meters squared (kg/m²). All measurements were performed with participants standing barefoot and in a fasting state. Blood pressure was measured using an automated sphygmomanometer (Omron HEM-907, Japan). Each participant rested in a seated position for at least 5 min before three blood pressure readings were taken at intervals of no less than 2 min. The average of the three readings was used as the final blood pressure value. Fasting blood samples were collected in the early morning and analyzed in the hospital’s clinical laboratory using an automated biochemical analyzer (Hitachi 7600, Japan). Biochemical indicators included fasting plasma glucose (FPG), lipid profiles including triglycerides (TG), total cholesterol (TC), high-density lipoprotein cholesterol (HDL-C), and low-density lipoprotein cholesterol (LDL-C), as well as liver function markers such as alanine aminotransferase (ALT), albumin (ALB), total bilirubin (TBIL), and serum globulin (GLB). Body composition analysis was conducted using a bioelectrical impedance analyzer (BIA AFE CS1253, ChinHai Technology, China). Nine parameters were measured, including body fat mass (FatM), fat-free mass (LeanM), muscle mass (Muscle), estimated bone mass (Bone), total body water (Water), body water percentage (Water%), extracellular water percentage (ECW%), visceral fat rating (VFR), and basal metabolic rate (BMR). All measurements were performed by the same operator using the same equipment according to standardized procedures to ensure consistency and reproducibility.

MAFLD diagnostic criteria

The diagnosis of MAFLD is based on international consensus criteria. Hepatic steatosis was assessed through abdominal ultrasonography using a high-resolution Doppler ultrasound device (Mindray DC-8, Mindray Biomedical Electronics Co., Ltd., Shenzhen, China). Participants were diagnosed with MAFLD if hepatic steatosis was present along with one of the following three conditions: (1) T2DM; (2) overweight or obesity, with BMI ≥ 23 kg/m² in Asian populations; or (3) at least two metabolic risk abnormalities, including: waist circumference(WC) ≥ 102 cm for men or ≥ 88 cm for women, blood pressure ≥ 130/85 mmHg or currently on antihypertensive treatment, prediabetes, plasma triglycerides ≥ 1.7 mmol/L or the use of lipid-lowering agents, high-density lipoprotein cholesterol (HDL-C) < 1.0 mmol/L for men or < 1.3 mmol/L for women, or on special treatments, homeostasis model assessment of insulin resistance (HOMA-IR) ≥ 2.5, C-reactive protein (hs-CRP) > 2 mg/L¹⁹.

Feature selection strategy

To ensure the scientific rigor of variable selection and the robustness of model construction, this study integrated statistical methods with clinical knowledge during the feature selection process, guided by the following three principles: initially, a total of 32 candidate variables were available from the health examination dataset. Based on the study objective of developing a non-invasive screening/diagnostic model for MAFLD, a preliminary selection was performed to retain variables that were directly obtainable without invasive procedures. Through this step, 13 non-invasive variables, including body composition indicators and basic demographic characteristics, were retained for further analysis. Next, multicollinearity was assessed using the variance inflation factor (VIF) in an iterative procedure: predictors with the highest VIF were sequentially removed, and VIFs were recalculated after each removal until all remaining variables satisfied the predefined threshold (VIF < 10). This cutoff is commonly used in regression-based and epidemiological studies, where VIF values > 10 are widely regarded as indicative of serious multicollinearity²⁰. Following multicollinearity control, SHapley Additive exPlanations (SHAP) derived from the Extreme Gradient Boosting (XGBoost) model were applied as a post-hoc interpretability tool to assess the relative importance and stability of the remaining features. In this study, VIF-based multicollinearity assessment served as a strict exclusion criterion, whereas SHAP was applied only within the VIF-filtered feature set to support interpretability and stability assessment, rather than to override VIF-based exclusion. Importantly, SHAP analyses were conducted exclusively on models trained using the training set²¹. After the training set was defined, the XGBoost model was trained and optimized within the training set using cross-validation, and SHAP values were subsequently computed on the training data only. The internal test set and the independent external validation cohort were not used for SHAP computation and did not participate in the feature selection process, thereby avoiding information leakage²². Through this sequential process, six variables were ultimately retained for final model construction, all of which satisfied the predefined multicollinearity criterion (VIF < 10; Table S3). This feature selection strategy ensured alignment with the study objectives while balancing statistical robustness, clinical relevance, and interpretability.

Machine learning modeling and external validation set evaluation

To comprehensively evaluate the performance of different machine learning methods in identifying MAFLD, this study built and compared eight mainstream classification models, including Logistic Regression, Decision Tree, Random Forest, Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Gradient Boosting Decision Tree (GBDT), LightGBM, and K-Nearest Neighbors (KNN)^23–25. The machine learning models were constructed based on the variables selected through the previous feature selection process. Prior to model construction, all continuous variables were standardized using z-score normalization. Importantly, the normalization parameters were calculated exclusively from the training set and subsequently applied to the internal test set and the external validation cohort to avoid information leakage. The development cohort (2017–2021) was randomly divided into a training set and an internal test set at a ratio of 7:3. The training set was used exclusively for model development, including model fitting, hyperparameter tuning, and model selection. Within the training set, tenfold cross-validation was applied for hyperparameter optimization and model selection. In each cross-validation iteration, models were trained on nine folds and validated on the remaining fold. Hyperparameters for each algorithm were tuned within predefined ranges using a cross-validation-based search strategy, and the optimal configuration was determined based on average cross-validation performance²⁶. After completion of cross-validation, a final model for each algorithm was retrained on the full training set using the optimized hyperparameters. These finalized models were then evaluated on the internal test set, which was not involved in any stage of model training or tuning, to obtain an unbiased estimate of internal performance. To further assess temporal generalizability, the finalized models trained on the full training set were subsequently applied to an independent external validation cohort collected between January 2022 and December 2023. This cohort did not participate in any stage of model training, hyperparameter optimization, or internal evaluation. Given the moderate class imbalance between MAFLD and non-MAFLD participants, no oversampling or undersampling techniques were applied. Instead, imbalance was addressed through appropriate model selection, adjustment of class weights where applicable, and the use of multiple complementary evaluation metrics, including AUC-ROC, recall, F1 score, and Brier score. Model interpretability was assessed using feature importance measures. For tree-based ensemble models (XGBoost, GBDT, and LightGBM), SHapley Additive exPlanations (SHAP) were applied as a post-hoc interpretability method based on the finalized models, enabling quantification of the marginal contribution of each feature without influencing model training or performance evaluation.

Statistical analysis

All statistical analyses were performed in a Python 3.11 environment, with a significance level set at p < 0.05. The study participants were divided into two groups based on the presence or absence of MAFLD: the case group and the control group. Descriptive statistics were first conducted to summarize the baseline characteristics of the two groups. Continuous variable were tested for normality. Variables that followed a normal distribution were expressed as mean ± standard deviation, and group comparisons were performed using independent t-tests. For non-normally distributed data, the median (interquartile range) [M(P25, P75)] was used, and comparisons were made using the Mann–Whitney U test. Categorical variables were expressed as frequencies and percentages [n(%)], and group comparisons were conducted using the chi-square test or Fisher’s exact test. To further explore the strength of the association between each variable and MAFLD, point-biserial correlation analysis was used to examine the relationship between continuous variables and the binary outcome (presence or absence of MAFLD), identifying potential important variables. Simultaneously, collinearity among candidate variables was assessed by calculating the VIF to identify multicollinearity issues. Variables with a VIF greater than 10 were excluded, while those with acceptable collinearity and strong representativeness were retained for subsequent modeling analysis. To evaluate whether body composition analysis results were independent risk factors for MAFLD, two-level logistic regression models were constructed. In Model 1, univariate logistic regression was performed for each body composition indicator to assess their association with MAFLD. The unadjusted odds ratios (OR) and 95% confidence intervals (95% CI) were calculated to identify potential significant predictors of MAFLD. Each variable was evaluated independently without considering the influence of other factors. In Model 2, multivariate logistic regression analysis was conducted, incorporating potential confounding factors such as gender, age, waist circumference, BMI, hypertension, diabetes, smoking, and alcohol consumption. This analysis allowed us to examine the independent association of each body composition indicator with MAFLD after adjusting for these confounding effects. To further validate the robustness and consistency of these variables across different populations, stratified regression analyses were conducted based on three key demographic features—gender (male vs. female), age (≥ 40 years vs. < 40 years), and BMI levels (< 23, 23–28, ≥ 28 kg/m²). The OR and 95% CI for each subgroup were calculated, systematically exploring the independent identification ability and applicability of body composition variables within each subgroup. This analysis not only provides a basis for variable selection in subsequent machine learning models but also validates the scientific value and clinical potential of body composition analysis in MAFLD risk assessment from a traditional statistical perspective.

Results

Population characteristics and correlation analysis

As shown in Table 1, a total of 23,348 participants undergoing health examinations were included in this study, with 9,479 (40.6%) classified as having MAFLD and 13,869 (59.4%) as non-MAFLD. Significant differences were observed between the MAFLD and non-MAFLD groups in terms of sex distribution, body weight, blood pressure, lipid levels, and body composition parameters (all P < 0.001). Specifically, the proportion of males in the MAFLD group was significantly higher than that in the non-MAFLD group (90.8% vs. 43.0%), and body weight was also markedly elevated (76.35 vs. 58.95 kg). In addition, the MAFLD group exhibited higher systolic blood pressure (127.00 mmHg) and diastolic blood pressure (79.00 mmHg) compared to the non-MAFLD group (116.00 mmHg and 71.00 mmHg, respectively). In terms of lipid profiles, TG (1.34 mmol/L) and TC (5.13 mmol/L) were notably elevated in the MAFLD group compared to the non-MAFLD group (1.15 mmol/L and 4.82 mmol/L, respectively). Body composition analysis indicated higher Fat% (14.60%) and VFR (6.00 kg) in the MAFLD group than in the non-MAFLD group (11.60% and 5.00 kg, respectively). These findings suggest that MAFLD is closely associated with male sex, obesity, hyperlipidemia, hypertension, and visceral adiposity, supporting its metabolic syndrome nature. Further correlation analysis (Fig. 3, Supplementary Table S1) demonstrated that VFR (r = 0.743), WC (r = 0.717), and BW (r = 0.677) showed the strongest positive correlations with MAFLD (all P < 0.001). Sex (r = − 0.485) was negatively correlated, indicating higher risk in males. Other variables such as FatM (r = 0.483), HDL-C (r = 0.439), and DBP (r = 0.358) also showed significant positive correlations. Smoking status (r = 0.278) and Water% (r = − 0.236) were significantly associated with MAFLD, with Water% showing a negative correlation (P < 0.001). Weaker but still significant correlations were observed for age (r = 0.137) and hypertension (r = 0.124; both P < 0.001). In contrast, Fat% and GLOB were not significantly associated with MAFLD (P = 0.172 and P = 0.84, respectively). These results highlight the critical role of BW, lipids, BP, and visceral fat in MAFLD development and their link to metabolic syndrome. Additionally, sex, smoking, and Water% may also contribute to MAFLD identification.

Table 1.

Characteristics of the total population and comparison between the MAFLD and non-MAFLD groups.

Variables	Total population (n = 23,348)	No-MAFLD (n = 13,869)	MAFLD (n = 9479)	P-value
Male,n(%)	14,562(62.37%)	5959 (43.0%)	8603 (90.8%)	< 0.001
Age,years	42.00 (34.00–51.00)	40.00 (32.00–49.00)	44.00 (36.00–52.00)	< 0.001
Height,cm	165.00 (159.00–170.60)	162.00 (156.80–168.40)	168.50 (164.00–172.85)	< 0.001
Weight,kg	66.15 (57.10–75.35)	58.95 (52.95–65.55)	76.35 (70.57–83.20)	< 0.001
SBP,mmHg	121.00 (111.00–131.00)	116.00 (107.00–127.00)	127.00 (118.00–136.00)	< 0.001
DBP,mmHg	74.00 (67.00–83.00)	71.00 (64.00–78.00)	79.00 (72.00–87.00)	< 0.001
WC,cm	84.00 (76.00–92.00)	78.00 (72.00–83.00)	93.00 (89.00–98.00)	< 0.001
BMI,kg/m2	24.27 (21.94–26.76)	22.47 (20.71–24.13)	27.02 (25.41–28.97)	< 0.001
ALT,U/L	23.00 (15.00–35.00)	18.00 (13.00–26.00)	32.00 (23.00–48.00)	< 0.001
TP,g/L	73.50 (70.80–76.40)	73.40 (70.60–76.20)	73.80 (71.10–76.70)	< 0.001
ALB,g/L	46.90 (45.00–48.70)	46.60 (44.80–48.50)	47.20 (45.30–49.00)	< 0.001
GLOB,g/L	26.60 (24.30–29.00)	26.60 (24.30–29.10)	26.60 (24.30–29.00)	0.845
FBG,mmol/L	5.29 (4.96–5.72)	5.18 (4.88–5.52)	5.52 (5.12–6.04)	< 0.001
TC,mmol/L	4.95 (4.38–5.59)	4.82 (4.27–5.46)	5.13 (4.56–5.78)	< 0.001
TG,mmol/L	1.47 (0.97–2.29)	1.15 (0.82–1.69)	2.08 (1.46–3.14)	< 0.001
HDL-C,mmol/L	1.24 (1.07–1.44)	1.34 (1.17–1.54)	1.12 (0.99–1.26)	< 0.001
LDL-C,mmol/L	2.83 (2.32–3.36)	2.78 (2.32–3.31)	2.89 (2.32–3.43)	< 0.001
FatM, kg	16.70 (13.40–20.70)	14.60 (11.75–17.95)	19.90 (16.75–24.30)	< 0.001
LeanM, kg	50.75 (39.95–57.10)	41.90 (37.85–51.70)	56.95 (52.65–61.00)	< 0.001
Muscle, kg	48.10 (37.60–54.15)	39.40 (35.70–49.00)	54.00 (49.90–57.85)	< 0.001
Bone, kg	2.70 (2.30–2.95)	2.40 (2.15–2.75)	2.95 (2.75–3.15)	< 0.001
Water, L	34.15 (28.60–38.70)	29.85 (26.75–34.95)	38.35 (35.10–41.70)	< 0.001
Water%	50.80 (48.10–53.90)	51.50 (48.80–54.70)	49.90 (47.00–52.70)	< 0.001
ECW%	14.20 (12.05–15.55)	12.55 (11.15–14.15)	15.60 (14.75–16.55)	< 0.001
VFR, level	9.00 (5.00–12.00)	6.00 (4.00–9.00)	13.00 (11.00–14.00)	< 0.001
BMR, kcal/day	1406.00 (1180.00–1596.00)	1230.00 (1114.00–1430.00)	1590.00 (1458.00–1723.50)	< 0.001
Hypertension,n(%)	1495 (6.4%)	542 (3.9%)	953 (10.1%)	< 0.001
Diabetes,n(%)	579 (2.48%)	195 (1.4%)	385 (4.1%)	< 0.001
Smoking,n(%)	8582 (36.76%)	3561 (25.7%)	5021 (53.0%)	< 0.001
Alcohol,n(%)	5412 (23.18%)	2578 (18.6%)	2835 (29.9%)	< 0.001

Continuous variables are presented as median (interquartile range, IQR), and categorical variables as number (percentage). Comparisons between groups were performed using the Mann–Whitney U test for continuous variables and the Chi-square test for categorical variables. A p-value < 0.05 was considered statistically significant.

BMI, body mass index; ALT, alanine aminotransferase; TP, total protein; ALB, albumin; GLOB, globulin; FBG, fasting blood glucose; TC, total cholesterol; TG, triglyceride; HDL-C, high-density lipoprotein cholesterol; LDL-C, low-density lipoprotein cholesterol; FatM, fat mass; LeanM, lean mass; ECW%, extracellular water percentage; VFR, visceral fat rating; BMR, basal metabolic rate; WC, waist circumference; SBP, systolic blood pressure; DBP, diastolic blood pressure; Water, total body water; Water%, body water percentage; Muscle, muscle mass; Bone,estimated bone mass.

Fig. 3.

Correlation Analysis of Variables Associated with MAFLD. Absolute Spearman correlation coefficients between clinical variables and MAFLD status. Variables are ranked by the strength of their absolute correlation. Bar colors indicate significance levels: significant positive (blue), significant negative (green), and not significant (gray) (p < 0.05). BMI, body mass index; ALT, alanine aminotransferase; TP, total protein; ALB, albumin; GLOB, globulin; FBG, fasting blood glucose; TC, total cholesterol; TG, triglyceride; HDL-C, high-density lipoprotein cholesterol; LDL-C, low-density lipoprotein cholesterol; FatM, fat mass; LeanM, lean mass; ECW%, extracellular water percentage; VFR, visceral fat rating; BMR, basal metabolic rate; WC, waist circumference; SBP, systolic blood pressure; DBP, diastolic blood pressure; TBIL, total bilirubin; Water, total body water; Water%, body water percentage; Muscle, muscle mass; Bone,estimated bone mass.

Feature selection

Based on the study objectives, a total of 13 variables were included for analysis, comprising nine body composition parameters, as well as BMI, WC, sex, and age. Multicollinearity among variables was assessed by calculating the VIF. As shown in Supplementary Table S2, certain variables—such as WC (VIF = 9.24), BMI (VIF = 20.80), fat mass (VIF = 38.99), and fat-free mass (VIF = 1542.75)—showed high VIF values, suggesting considerable collinearity. To reduce the impact of multicollinearity, variables with excessively high VIFs were excluded, and core features with clinical relevance and acceptable VIF levels were retained. Final VIF statistics are presented in Supplementary Table S3. Further selection was conducted using the XGBoost model combined with SHAP analysis. As illustrated in Fig. 4, SHAP values and corresponding bar plots indicate the relative importance of each feature in predicting MAFLD (Supplementary Table S4). VFR, WC, and weight had the highest SHAP contributions, with VFR and WC ranking among the most influential predictors. Integrating SHAP results, multicollinearity evaluation, and clinical relevance, the final set of features used for model construction included VFR, WC, BMI, weight, ECW%, and total body water. All selected variables had VIF values below the predefined threshold, ensuring model stability and predictive performance. Detailed results and the final list of selected features are provided in the supplementary materials.

Fig. 4.

Feature Importance in MAFLD Prediction Based on SHAP Analysis with XGBoost. (A) SHAP summary plot showing the distribution and direction of each feature’s impact on the model output. Each point represents a single observation, with colors indicating the feature value (red = high, blue = low). Features are ranked by importance, with VFR, WC, and BMI contributing the most to MAFLD prediction. (B) Mean absolute SHAP values for each feature, indicating their overall impact on model predictions. VFR and WC exhibited the highest mean SHAP values, followed by BMI and ECW%. The final model incorporated features with high predictive value, clinical relevance, and acceptable multicollinearity levels. VFR, visceral fat rating; WC, waist circumference; BMI, body mass index; ECW%, extracellular water percentage; Water%, body water percentage; FatM, fat mass; BMR, basal metabolic rate; LeanM, lean mass; Water, total body water; Muscle, muscle mass; Bone,estimated bone mass.

Model performance and feature importance

Table 2 presents the performance metrics of eight classification models in identifying MAFLD. Among them, GBDT, LightGBM, and XGBoost exhibited the highest AUC values (all 0.967), indicating excellent discriminative ability. These models also demonstrated relatively high accuracy (0.916), recall (ranging from 0.962 to 0.966), and F1 scores (0.903), with the lowest Brier scores (0.063), reflecting good overall calibration. The Random Forest model achieved the highest recall (0.969) and maintained competitive performance across other metrics. In contrast, Logistic Regression and SVM showed slightly lower values in recall and F1 score, suggesting a moderate decrease in sensitivity and balance. The KNN model showed the lowest AUC (0.956) and the highest Brier score (0.072), indicating relatively weaker predictive performance. Overall, tree-based ensemble models (GBDT, LightGBM, XGBoost, and Random Forest) demonstrated superior and stable performance across all evaluation criteria. Feature importance results for all models are presented in Fig. 5. Variables such as VFR, WC, and BMI consistently ranked high across different algorithms, with VFR and WC identified as the top contributors in models such as XGBoost, GBDT, and LightGBM. In addition, SHAP analysis further confirmed that VFR, WC, and ECW% had the greatest impact on model output, with VFR and WC showing the highest SHAP values.

Table 2.

Performance comparison of eight classification models in MAFLD identification.

Model	AUC	Accuracy	Recall	F1 Score	Brier Score
Decision Tree	0.963	0.914	0.959	0.901	0.066
GBDT	0.967	0.916	0.966	0.903	0.063
KNN	0.956	0.905	0.931	0.888	0.072
LightGBM	0.967	0.916	0.962	0.903	0.063
Logistic Regression	0.964	0.899	0.876	0.875	0.073
Random Forest	0.967	0.914	0.969	0.902	0.064
SVM	0.963	0.902	0.897	0.881	0.073
XGBoost	0.967	0.916	0.963	0.903	0.063

This table summarizes the predictive performance of eight machine learning models based on five evaluation metrics: area under the curve (AUC), accuracy, recall, F1 score, and Brier score. Lower Brier scores indicate better calibration performance.

Fig. 5.

Feature Importance of Predictive Variables Across Multiple Machine Learning Models. (A)–(C) Mean SHAP values of top features in XGBoost, GBDT, and LightGBM models, indicating the average magnitude of each feature’s impact on MAFLD prediction. (D)–(F) SHAP summary plots for XGBoost, GBDT, and LightGBM, showing both the magnitude and direction of feature impacts, with color representing feature value (red = high, blue = low). (G–K) Feature importance derived from traditional methods: (G) Decision Tree, (H) KNN (permutation importance), (I) SVM (absolute coefficients), (J) Random Forest, and (K) Logistic Regression (absolute coefficients). Across models, VFR and WC consistently ranked as top predictors, followed by ECW%, age, and sex. Abbreviations: VFR, visceral fat rating; WC, waist circumference; ECW%, extracellular water percentage; BMR, basal metabolic rate; SHAP, Shapley additive explanations; KNN, k-nearest neighbors; SVM, support vector machine.

External validation performance

The performance of eight machine learning models was evaluated using an external validation dataset. As shown in Table 3, all models demonstrated stable predictive performance across multiple metrics, including AUC, accuracy, recall, and F1 score. The XGBoost model achieved the highest AUC (0.955), followed by Random Forest (0.954) and LightGBM (0.953). In terms of accuracy, XGBoost (0.894) slightly outperformed LightGBM (0.894) and GBDT (0.899). For recall, the Random Forest model showed the highest value (0.927), followed by XGBoost (0.923) and Decision Tree (0.916), indicating strong sensitivity in identifying positive cases. Regarding the F1 score, XGBoost (0.889) showed marginally higher values compared to GBDT (0.887), reflecting a favorable balance between precision and recall. Brier scores were close to 0 for all models, indicating good calibration. The Decision Tree and KNN models yielded Brier scores of 0.088 and 0.090, respectively. Overall, the models maintained consistent performance on the external dataset, suggesting good generalizability across different data distributions.

Table 3.

External Validation Performance of Eight Classification Models for MAFLD Identification.

Model	AUC	Accuracy	Recall	F1 Score	Brier Score
Decision Tree	0.944	0.887	0.916	0.882	0.087
GBDT	0.952	0.899	0.923	0.887	0.085
KNN	0.934	0.886	0.895	0.878	0.090
LightGBM	0.953	0.894	0.921	0.889	0.084
Logistic Regression	0.951	0.876	0.841	0.862	0.088
Random Forest	0.954	0.891	0.927	0.886	0.083
SVM	0.951	0.881	0.864	0.869	0.088
XGBoost	0.955	0.894	0.923	0.889	0.083

This table summarizes the performance of eight machine learning models evaluated on an external validation dataset. Metrics include area under the curve (AUC), accuracy, recall, F1 score, and Brier score. The XGBoost model achieved the highest AUC (0.955) and maintained a balanced performance across other metrics. Random Forest showed the highest recall (0.927), while LightGBM and GBDT also demonstrated competitive predictive performance. All models yielded low Brier scores, indicating good calibration and model reliability on unseen data.

Logistic regression analysis of body composition variables associated with MAFLD in the overall and stratified populations

As shown in Table 4, logistic regression analyses revealed that body composition indicators were significantly associated with MAFLD. In Model 1, all variables, including Fat% (OR = 1.039), FatM (OR = 1.248), and LeanM (OR = 1.187), were significantly related to MAFLD (P < 0.001). After adjustment in Model 2, Fat% (OR = 1.115), FatM (OR = 1.134), and VFR (OR = 1.547) remained significantly associated with MAFLD, while LeanM showed a negative association (OR = 0.976). ECW% was marginally non-significant (P = 0.052).

Table 4.

Logistic regression analysis of body composition indicators associated with MAFLD in the total population.

Variables	Model1		Model2
	OR(95%CI)	P-value	OR(95%CI)	P-value
FatM	1.248(1.240–1.257)	< 0.001	1.134(1.111–1.574)	< 0.001
LeanM	1.187(1.182–1.192)	< 0.001	0.976(0.965–0.988)	< 0.001
Muscle	1.195(1.190–1.201)	< 0.001	0.975(0.963–0.989)	< 0.001
Bone	54.734(49.170–60.927)	< 0.001	0.721(0.579–0.899)	< 0.001
Water	1.289(1.281–1.298)	< 0.001	0.950(0.937–0.963)	< 0.001
Water%	0.892(0.886–0.897)	< 0.001	0.900(0.888–0.913)	< 0.001
ECW%	2.775(2.702–2.850)	< 0.001	0.962(0.926–1.000)	0.052
VFR	2.270(2.221–2.319)	< 0.001	1.547(1.466–1.632)	< 0.001
BMR	1.006(1.006–1.006)	< 0.001	0.999(0.999–0.999)	< 0.001

Model 1 presents univariate logistic regression results for each body composition variable. Model 2 presents multivariate logistic regression results adjusted for age, sex, BMI, WC, hypertension, diabetes, smoking, and alcohol consumption. ORs with 95% CI and p-values are reported. VFR and BMR remained significant predictors after adjustment. P-values < 0.05 were considered statistically significant.

FatM, fat mass; LeanM, lean mass; Muscle, muscle mass; Bone, bone mass; Water, total body water; Water%, percentage of total body water; ECW%, extracellular water percentage; VFR, visceral adipose rating; BMR, basal metabolic rate; OR, odds ratio; CI, confidence interval.

Sex-stratified logistic regression analyses were conducted to examine the associations between body composition indicators and MAFLD (Supplementary Table S5–6). In Model 1, all indicators were significantly associated with MAFLD in both sexes. In Model 2, VFR remained strongly associated with MAFLD in both males (OR = 1.564, 95% CI: 1.510–1.620) and females (OR = 1.787, 95% CI: 1.627–1.964). BMR also remained significant in both sexes, while Water% showed consistent inverse associations (male: OR = 0.928; female: OR = 0.874). After adjustment, some indicators such as LeanM, Muscle, and Bone retained significance only in males.

Age-stratified logistic regression analyses were also conducted (< 40 vs. ≥ 40 years; Supplementary Table S7-8). In unadjusted Model 1, all indicators were significantly associated with MAFLD in both age groups. In the adjusted Model 2, VFR remained a robust predictor in both subgroups (≥ 40 years: OR = 1.331, 95% CI: 1.271–1.392; < 40 years: OR = 2.137, 95% CI: 1.939–2.356). BMR retained significance in both groups, while Water% showed consistent inverse associations. Some indicators, including bone mass (≥ 40 years: OR = 0.599; < 40 years: OR = 0.422) and lean mass (< 40 years: OR = 0.954), remained significantly associated, whereas others such as ECW% lost statistical significance in the older group after adjustment (p = 0.583).

BMI-stratified logistic regression analyses were further performed across three subgroups (< 24, 24–28, and > 28 kg/m²; Supplementary Table S9–10). In Model 1, all body composition indicators were significantly associated with MAFLD. After adjustment in Model 2, VFR consistently showed strong associations across all BMI strata (< 24: OR = 2.305, 95% CI: 2.142–2.483; 24–28: OR = 1.444, 95% CI: 1.381–1.501; > 28: OR = 1.178, 95% CI: 1.069–1.298). BMR remained significant in all subgroups. ECW% was positively associated in the BMI > 28 group (OR = 1.008, 95% CI: 1.003–1.105) but not in the 24–28 kg/m² group (p = 0.853). Other indicators, such as lean mass, muscle mass, and bone mass, were significant in the lower BMI groups but lost or attenuated in significance among participants with BMI > 28 kg/m².

Discussion

This study utilized real-world data from individuals undergoing routine health examinations at the Health Management Center of the Third Xiangya Hospital, Central South University, to systematically evaluate the association between body composition parameters and the risk of metabolic dysfunction–associated MAFLD. A total of nine body composition indicators derived from BIA were initially considered. After accounting for multicollinearity and ensuring clinical interpretability, six key predictors were ultimately selected based on SHAP feature importance rankings from the XGBoost model: VFR, WC, body weight, BMI, total body water, and extracellular water percentage. Using these variables, eight commonly used machine learning algorithms were developed. Among them, XGBoost and GBDT demonstrated the highest discriminative performance. The models demonstrated strong discriminative performance, with area under the curve (AUC) values exceeding 0.96 in both internal cross-validation and independent external validation cohorts, indicating strong generalizability.

Building upon the selected features and modeling framework, we further examined the specific associations between individual body composition indicators and MAFLD. The results showed that individuals with MAFLD exhibited significantly higher levels of metabolic parameters, including BMI, waist circumference, blood pressure, triglycerides, and VFR, compared to those without the disease, which is consistent with previous findings highlighting the clustering of metabolic abnormalities in MAFLD populations^27–29. Among all evaluated indicators, VFR consistently emerged as the most stable and independent predictor across all models and subgroups, emphasizing its central role in the pathogenesis of MAFLD. Visceral fat, in contrast to subcutaneous fat, is metabolically active and contributes to disease progression by secreting pro-inflammatory cytokines such as TNF-α and IL-6, as well as free fatty acids that are delivered directly to the liver through the portal circulation^30–32. These mediators promote hepatic insulin resistance, lipid accumulation, and oxidative stress, thereby facilitating the onset and advancement of MAFLD^33–35. Although VFR is an indirect estimate derived from BIA based on waist circumference, age, and body weight, it offers greater physiological relevance than conventional indicators such as BMI, as it more accurately reflects abdominal fat distribution³⁶. This helps address the limitations of traditional obesity metrics in identifying metabolic risk. In addition to VFR, elevated ECW% was significantly associated with MAFLD, potentially indicating extracellular fluid retention linked to chronic low-grade inflammation³⁷. This may reflect increased capillary permeability and interstitial fluid accumulation, commonly seen in metabolic disorders. Higher ECW% may also be related to impaired renal function, poor nutritional status, or reduced muscle mass—all of which could contribute to MAFLD progression^38,39. Thus, ECW% may serve as a useful indicator of fluid imbalance, systemic inflammation, and metabolic dysfunction in clinical risk assessment⁴⁰. Total body water was consistently inversely associated with MAFLD, potentially reflecting better fluid homeostasis or greater muscle mass, both of which may confer a protective effect^41,42. Overall, these indicators represent distinct dimensions of MAFLD pathophysiology, including fat distribution, fluid balance, and basal metabolism. This underscores that MAFLD is not merely a consequence of general obesity but rather a complex outcome of systemic metabolic dysregulation.

Stratified analyses further revealed heterogeneity in the effects of body composition variables across different populations. These subgroup-specific distributions (Supplementary Tables S5, S7, and S9) indicate notable sex, age, and BMI-related differences in visceral fat, fluid balance, and skeletal composition, which may underlie variations in MAFLD risk. In sex-stratified models, lean mass, muscle mass, and bone mass remained significantly associated with MAFLD in males but not in females after multivariable adjustment, suggesting modulation by sex-specific factors such as hormonal status and fat distribution patterns^43–45. Age-stratified analyses showed relatively modest differences, with VFR, BMR, and total body water remaining consistently associated with MAFLD across age groups, while ECW% and lean mass demonstrated stronger effects in younger individuals. Notably, in individuals with BMI > 28 kg/m², many indicators lost statistical significance, with only VFR and ECW% remaining robustly associated with MAFLD, possibly reflecting collinearity or a ceiling effect in obese populations. These findings underscore the particularly strong discriminative value of VFR in high-BMI groups^46,47.

To contextualize these findings, we compared our results with previously reported non-invasive models for fatty liver disease or MAFLD identification. Reported discrimination performance varies by data modality and reference standard; commonly used clinical or laboratory-based scores typically show AUC values around 0.76–0.83 in meta-analytic evidence, with some population-specific validations reporting AUCs of approximately 0.82–0.91. In the present study, machine learning models based on BIA-derived body composition indicators achieved AUC values exceeding 0.96 in both internal evaluation and an independent temporal validation cohort. Differences in features, study populations, and outcome definitions should therefore be considered when interpreting cross-study comparisons.

Compared with previous studies that primarily relied on conventional metabolic indicators such as BMI, WC, ALT, and TG^48,49, this study expands the feature space by incorporating multiple BIA-derived body composition indicators, including VFR, extracellular water percentage, and total body water, capturing complementary physiological dimensions of fat distribution, fluid balance, and basal metabolism. We combined multicollinearity diagnostics with SHAP-based explainability to support interpretable feature selection and developed multiple machine learning models—particularly tree-based algorithms such as XGBoost—showing consistently strong discrimination and robust temporal generalizability. From a practical perspective, BIA is a non-invasive, portable, and low-cost technique suitable for routine health examinations and primary care settings^50,51. When integrated with machine learning, this approach provides an interpretable and scalable framework for MAFLD screening and classification, supporting early identification and stratified management in real-world populations. Recent studies have increasingly adopted explainable machine learning frameworks that integrate tree-based models with SHAP-based interpretation to derive clinically meaningful predictors in metabolic and endocrine research. For example, explainable machine learning approaches have been applied to the identification of metabolic dysfunction–associated steatotic liver disease and to the prediction of endocrine, nutritional, and metabolic mortality, demonstrating the value of combining model interpretability with robust feature selection strategies^52,53. Our analytical framework is consistent with this emerging paradigm, while extending it to a large health examination cohort and focusing on non-invasive, BIA-derived body composition indicators for MAFLD identification.

Despite the strengths of this study, including a large sample size, external validation, methodological diversity, and strong model interpretability, several limitations should be acknowledged. First, the cross-sectional design precludes causal inference, and future longitudinal studies are needed to validate the predictive value of the model in disease progression. Second, although the BIA-estimated VFR is widely used in clinical practice, its accuracy is inferior to imaging-based gold standards such as CT or MRI. Future studies should incorporate high-resolution imaging data to further assess the model’s precision. Third, as the data were derived from a single-center health examination population, the geographic and ethnic generalizability is limited. Although the external validation cohort was drawn from a different time period, the model has not yet been tested in multi-center or multi-ethnic populations. Fourth, detailed lifestyle and behavioral factors (e.g., diet, physical activity, alcohol consumption, and sleep) were not available in the health examination dataset. This may lead to residual confounding and also limits prevention-oriented interpretation, including the use of actionable explainability approaches to identify modifiable intervention targets. Future prospective studies that collect lifestyle exposures and follow participants over time will be necessary to enable intervention-focused interpretability and to better support prevention strategies. Future research should integrate multi-omics data—such as metabolomics, genomics, and gut microbiota—to develop a multimodal prediction framework and promote the integration of such models into clinical workflows, ultimately enabling intelligent screening and personalized intervention strategies.

Conclusions

In conclusion, this study developed and evaluated predictive models for MAFLD based on BIA-derived body composition indicators in a large health examination cohort. VFR consistently showed the strongest and most stable association with MAFLD across all models and subgroups, suggesting its potential importance in metabolic risk assessment. Other indicators, including ECW%, BMR, and total body water, were also independently associated with MAFLD, reflecting potential disruptions in fluid balance and metabolic activity. Stratified analyses demonstrated consistent patterns across sex, age, and BMI categories, while also revealing subgroup-specific variations. The proposed models exhibited high predictive performance and strong generalizability, supporting their potential utility in clinical or preventive health settings to facilitate early screening and individualized metabolic risk stratification.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

MAFLD

Metabolic dysfunction-associated fatty liver disease

NAFLD

Nonalcoholic fatty liver disease

T2DM

Type 2 diabetes mellitus

BIA

Bioelectrical impedance analysis

VFR

Visceral fat rating

BMI

Body mass index

WC

Waist circumference

ALT

Alanine aminotransferase

ALB

Albumin

TBIL

Total bilirubin

GLB

Serum globulin

FPG

Fasting plasma glucose

TG

Triglycerides

TC

Total cholesterol

HDL-C

High-density lipoprotein cholesterol

LDL-C

Low-density lipoprotein cholesterol

ECW%

Extracellular water ratio

FatM

Fat mass

LeanM

Lean mass

Water

Total body water

Water%

Body water percentage

Muscle

Muscle mass

Bone

Estimated bone mass

BMR

Basal metabolic rate

SHAP

SHapley Additive exPlanations

VIF

Variance inflation factor

AUC

Area under the receiver operating characteristic curve

ROC

Receiver operating characteristic

SVM

Support vector machine

GBDT

Gradient boosting decision tree

KNN

K-Nearest Neighbors

Author contributions

YH: Performed data analysis, constructed models, and drafted the manuscript. YC: Assisted with data analysis, model development, and manuscript preparation. ZC: Contributed to machine learning model validation. RX: Supported external data processing and validation. FW: Conceived and supervised the study, interpreted the results, and critically revised the manuscript.

Funding

This study was supported by the following funding sources: the Natural Science Foundation of Hunan Province (Grant No. 2024JJ5520), the Changsha Municipal Natural Science Foundation (Grant No. kq2403054), the Hunan Provincial Program for Young Key Teachers in Universities (Grant No. 20240101–20261230).

Data availability

Due to privacy and ethical restrictions associated with the hospital-based health examination data, the raw data are not publicly available. The code used for data preprocessing, model training, evaluation, and interpretation is available at: https://github.com/hyaxuan23-lab/ML-for-Non-Invasive-MAFLD-Identification. Additional documentation is provided in the Supplementary Materials.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

This study was approved by the Ethics Committee of The Third Xiangya Hospital, Central South University (Approval Number: 225546). As this was a retrospective study using anonymized health examination data, the requirement for informed consent was waived by the ethics committee.

Consent for publication

Not applicable.