Machine learning-based identification of biochemical markers to predict hepatic steatosis in patients at high metabolic risk

Abstract

BACKGROUND

Metabolic-associated fatty liver disease (MAFLD) is the most common cause of chronic liver disease and remains under-recognized within the health check-up population. Ultrasonography during physical examination fail to accurately identify at-risk patients as they involve multiple metabolic aspects.

AIM

To rapidly identify hepatic steatosis patients from high-metabolic-risk populations and reduce medical costs.

METHODS

We analyzed all data from a prospective cohort study to identify potential predictors of MAFLD risk. The LASSO and recursive feature elimination were used to screen for feature selection. Four machine learning models were employed to construct the prediction model for hepatic steatosis.

RESULTS

We found that 86.2% of the 1011 individuals in the trial phase exhibited metabolic abnormalities, with 70.8% presenting with hepatic steatosis. After data cleaning, 711 participants (207 non-MAFLD patients vs 504 MAFLD patients) were included, and the prediction models were validated. After overlapping and reducing the feature set based on feature importance ranking, we developed an interpretable final XGBoost model with 10 features, achieving an area under the curve of 0.82.

CONCLUSION

We have introduced a valuable noninvasive tool for efficiently identifying hepatic steatosis patients in high-metabolic-risk populations. This tool may improve screening effectiveness and reduce medical costs.

Key Words: Metabolic-associated fatty liver disease; Machine learning; Prediction model; Hepatic steatosis; High metabolic risk population

Core Tip: We used a prospective cohort to develop and optimize a high-performance machine learning model, demonstrating its potential to screen the hepatic fat deposition in high-risk populations. We also integrate the facial and tongue diagnosis of traditional Chinese medicine (TCM) with the heterogeneity of metabolic-associated fatty liver disease (MAFLD) and introduce TCM-related indicators to increase the diversity of the metrics. Our model targets a more specific population and is applicable to a broader range of scenarios, which lays the foundation for significantly improving MAFLD check-up efficiency and reducing related medical expenses.

INTRODUCTION

Metabolic-associated fatty liver disease (MAFLD) is a chronic liver condition closely related to metabolic dysfunction[1]. The definition of MAFLD emphasizes the pathogenic role of metabolic dysfunction in disease progression and its systemic extrahepatic complications[2]. Despite the fact that fewer than 10% of MAFLD patients experience severe outcomes, such as cirrhotic complications and hepatocellular carcinoma, within 10-20 years after diagnosis[3], the absolute numbers of these cases are substantial due to the high prevalence of MAFLD. Other studies suggest that MAFLD has a greater risk of cardiovascular disease mortality than liver-related factors[4]. Therefore, prompt detection of high-risk individuals in clinical practice is imperative to enable early treatment initiation, which may contribute to optimized disease management and improved patient outcomes. The current guidelines suggest that high-metabolic-risk patients who are at risk for MAFLD should routinely undergo blood biochemical analysis and noninvasive assessment of hepatic steatosis and fibrosis[5]. In the early stages of hepatic steatosis, the normal range of transaminase levels and absence of discomfort symptoms make early identification difficult[6], and while traditional liver biopsies are used to characterize histological features, their invasive nature limits their use in widespread disease assessment and monitoring of therapy response. Magnetic resonance imaging-derived proton density fat fraction (MRI-PDFF) can objectively assess the fat content of the entire liver, but its high detection cost and poor accessibility limit its widespread application[7,8]. Ultrasound, which is widely used in health check-up populations for diagnosing hepatic steatosis, is less sensitive when the fat content is low, and its accuracy is also affected by the subjectivity of the examiner[9,10]. Liver stiffness quantification via vibration-controlled transient elastography, which is commercially implemented as FibroScan^® technology, employs ultrasound-derived signal attenuation analysis to calculate the controlled attenuation parameter (CAP), providing noninvasive hepatic steatosis assessment[11]. CAP has higher accuracy than conventional ultrasound examination[11]. However, its inadequate adoption rate during health check-ups limits further performance enhancement. Although the detection of fatty liver has been included in routine physical examinations, there are still deficiencies in efficiently and accurately identifying high-risk groups, and the awareness rate among the health check-up population remains low. Machine learning (ML) currently shows the potential to develop MAFLD prediction models based on clinical, biomarker, or imaging data. ML models based on ultrasound (including Liver Ultrasound Transient Elastography) and clinical data can reliably detect MAFLD and its complications, reducing diagnostic costs and the need for invasive liver biopsies[12]. However, MAFLD exhibits heterogeneity, which means that it may manifest with different characteristics and varying degrees of severity on an individual-to-individual basis[13]. The heterogeneity of MAFLD is reflected in its multisystem pathogenesis, indicating that it affects not only the liver but also other systemic diseases, such as metabolic syndrome, type 2 diabetes mellitus (T2DM), and cardiovascular diseases[1]. The multisystem involvement highlights the need for comprehensive consideration of various indicators in diagnosis and treatment. Nevertheless, the current application of ML in MAFLD prediction models has not yet fully embraced the comprehensiveness of multidimensional clinical indicators, and most are based on retrospective studies[12]. This is particularly concerning in the context of the heterogeneous and multifaceted nature of the health check-up population, which may prevent the models from capturing data characteristics from multiple perspectives, leaving the accuracy and reliability of the models open to question[14,15]. Traditional Chinese medicine (TCM) has been increasingly acknowledged and utilized globally over the past hundred years[16]. TCM adheres to the principle that “a great physician treats before an illness”. The multisystem characteristics of MAFLD align with the holistic perspective of TCM. The constitution reflects the overall state of the human body and is closely related to the onset and progression of diseases[17]. Some studies have reported that phlegm-dampness constitution is independently associated with MAFLD, and the increase in qi-stagnation and phlegm-dampness constitutions is closely related to T2DM in MAFLD patients[18,19]. In addition to laboratory examinations and anthropometric metrics, TCM constitutions can more effectively reflect the heterogeneity of diseases. Accordingly, we aimed to develop a noninvasive tool for rapidly and accurately identifying hepatic steatosis in health check-up populations, especially those at high metabolic risk. We attempted to identify risk factors for MAFLD by prospective cohort studies and develop a risk prediction model by integrating multidimensional clinical indicators, including TCM constitutions. This study will help promote the large-scale application of noninvasive models in fatty liver disease screening and is highly important for enhancing awareness of preventing and treating MAFLD.

MATERIALS AND METHODS

Study population

A prospective cohort study was conducted in a physical examination population in China to derivate and validate the hepatic steatosis prediction model. The derivation cohort comprised individuals who underwent physical examinations at the Health Checkup Center of Changhai Hospital in Shanghai between November 2024 and January 2025. This study adhered to the principles of the Helsinki Declaration. Each participant provided informed consent and willingly took part. The privacy of patient data was rigorously protected, and no interventions were conducted on the participants in this study.

Inclusion and exclusion criteria

Inclusion criteria: Participants had to be 18-65 years old and have received no treatment. The diagnostic criteria for MAFLD include the presence of hepatic steatosis confirmed by radiological imaging and at least one metabolic risk factor: Excessive overweight or obesity, T2DM, or metabolic dysregulation[20]. Hepatic steatosis was determined by ultrasonic attenuation (CAP) of the echo wave. We defined metabolic abnormalities as the presence of one or more of the followings: Excessive overweight or obesity, T2DM, or metabolic dysregulation.

Exclusion criteria: (1) Pregnant or lactating women; (2) Patients with malignant tumors or autoimmune liver disease; (3) Patients with viral hepatitis, autoimmune hepatitis, primary biliary cholangitis, or any other liver disease; and (4) Patients whose data were incomplete or missing.

Data collection and processing

We used all participants’ laboratory examinations, anthropometric metrics, demographic characteristics, and TCM constitution features to identify key features and construct predictive models. Consistent with established statistical protocols, covariates with missing value rates exceeding 30% underwent listwise deletion to preserve the validity of subsequent inferential analyses under the Missing Completely at Random assumption, whereas those with less than 30% missing values, median imputation was employed to handle the missing values. Additionally, considering the overall size of the dataset, we opted to use the synthetic minority oversampling technique for oversampling to address the issue of the insufficient number of non-MAFLD groups as much as possible.

TCM-related indicators collection process

The Traditional Chinese Medicine Tongue, Face, and Pulse Information Collection System for Constitution Identification, also known as the Intelligent Constitution Identifier (ICI), is provided by Shanghai Daosheng Medical Technology Co., Ltd. The volunteers were instructed to sit in front of the equipment, positioning their chins on designated supports. They were directed to maintain a neutral facial expression and extend their tongues, allowing for the recording of tongue and facial images. Concurrently, pulse information was obtained using a pulse diagnostic device. The equipment autonomously analyzed the data to generate a constitution report. To ensure the accuracy and reliability of the results obtained from the ICI, we arranged for two professional TCM physicians to conduct a rigorous review and assessment of the output results following the instrument detection. This process was designed to compensate for any potential deficiencies in the instrument’s algorithm. By incorporating manual verification and integrating the professional judgment of TCM physicians, we further enhanced the precision and reliability of the results, thereby providing accurate and reliable TCM constitution analysis for the subjects.

FibroScan CAP

All participants underwent CAP score determination using FibroScan (Handy, China) after enrolment. The procedure was strictly carried out by professionals in accordance with the standards for transient elastography data acquisition. CAP was considered valid only when the interquartile range (IQR) was < 40 dB/m. Furthermore, the CAP score acted as a predictive indicator for hepatic steatosis. Individuals were classified as having hepatic steatosis if their CAP values were ≥ 238 dB/m[21].

Model selection

In this study, four models were carefully chosen to represent distinct algorithmic paradigms and to ensure a comprehensive evaluation suitable for our structured clinical dataset. Specifically, XGBoost, an ensemble-based boosting method, was selected for its strong performance on structured tabular data and its ability to capture complex feature interactions. Random Forest (RF), an ensemble-based bagging method, was included due to its robustness against overfitting and its suitability for small-to-medium-sized datasets. Support Vector Machine (SVM), a maximum-margin classifier, was selected for its effectiveness in binary classification tasks involving high-dimensional data. Logistic regression, as a generalized linear model, was incorporated as a baseline given its simplicity, interpretability, and widespread clinical acceptance. This selection strategy allowed for a balanced comparison across tree-based, kernel-based, and linear modeling approaches.

Statistical analysis

Recursive feature elimination (RFE) recursively removes features and builds models to ultimately identify the most critical subset of features. RFE is commonly used for dimensionality reduction and improving model performance while reducing model complexity. LASSO regression achieves variable selection and sparsity constraints by incorporating an L1 regularisation term into the objective function that minimizes the sum of squared residuals. For comparative model evaluation, we implemented rigorous discrimination analysis through receiver operating characteristic (ROC) curve construction, with the integrated area under the curve (AUC) serving as the principal performance metric for distinguishing non-risk from high-risk hepatic steatosis cohorts. Continuous variables are summarized as the means ± SD, and non-normally distributed continuous variables are presented as medians with IQRs. Categorical variables are presented as counts with percentages. Hypothesis testing employed parametric Student's t tests or nonparametric Mann-Whitney tests for continuous measures, with Pearson χ² tests or Fisher's exact tests for categorical contrasts. A two-tailed P value of less than 0.05 was considered statistically significant. In the regression analysis, we conducted collinearity diagnostics on the independent variables to ensure the stability and reliability of the model. Collinearity was assessed using tolerance and the variance inflation factor (VIF). Generally, a tolerance value of less than 0.1 or a VIF greater than 10 indicates a serious collinearity problem.

Hyperparameter optimization was conducted to enhance model performance and ensure generalizability. For all ML models, a grid search strategy combined with five-fold cross-validation was employed. In the XGBoost model, hyperparameters such as the number of estimators, maximum tree depth, learning rate, minimum child weight, subsample ratio of training instances, subsample ratio of columns, and regularization term were tuned. For the RF model, the number of trees, maximum tree depth, minimum samples required to split an internal node, and minimum samples required at a leaf node were adjusted. In the Support Vector Machine model, the penalty parameter, kernel type, and kernel coefficient were optimized, while for Logistic regression, the regularization method and regularization strength were selected. The tuning process systematically explored a predefined range of values, and the optimal set of hyperparameters for each model was determined based on the highest mean AUC obtained from cross-validation. All hyperparameter tuning and model training procedures were implemented using the Scikit-learn and XGBoost libraries in Python.

RESULTS

Clinical characteristics

Initially, we gathered a group of 1011 individuals for the entire trial phase. The dataset contained 173 multidimensional clinical indicators detailed in Supplementary Table 1. As shown in Figure 1, 86.2% of the 1011 individuals presented metabolic abnormalities, including 742 overweight or obese individuals, 6 with dysglycaemia, and 34 with more than two metabolic risk factors. Among those individuals, 70.8% had hepatic steatosis, according to FibroScan analysis. As a result, we divided the subjects into non-MAFLD and MAFLD groups according to the diagnostic criteria.

Figure 1 Flow chart of the study design. TCM: Traditional Chinese medicine.

We then conducted intergroup comparisons using 173 indicators. Spearman correlation analysis (P < 0.01) revealed that 81 factors were significantly associated with MAFLD (Table 1). Among these, the top 10 indicators with the highest correlations were alanine aminotransferase (ALT), the ratio of aspartate aminotransferase (AST) to ALT, gamma-glutamyl transferase (γ-GT), triglyceride (TG), body mass index (BMI), AST, waist circumference (WC), body weight (BW), the ratio of uric acid (UA) to creatinine (Cr), and the ratio of Cr to BW. Notably, AST/ALT and Cr/BW were negatively correlated with the target variable. Finally, we performed collinearity diagnostics using regression analysis, employing the VIF to assess multicollinearity and minimize the mutual interference among similar indicators. After eliminating multicollinearity, we identified 16 indicators that retained high linear correlations without evidence of multicollinearity (Table 2).

Table 1 The 81 correlated factors associated with metabolic-associated fatty liver disease.

Spearman rank correlation
Indicators	r value	n	P value	Correlation coefficient	Correlation ranking
ALT	0.506	979	0.000	0.506	1
AST/ALT	-0.495	831	0.000	0.495	2
γ-GT	0.366	828	0.000	0.366	3
TG	0.358	917	0.000	0.358	4
BMI	0.353	867	0.000	0.353	5
AST	0.346	832	0.000	0.346	6
WC	0.331	774	0.000	0.331	7
Weight	0.303	778	0.000	0.303	8
UA/Cr ratio	0.288	972	0.000	0.288	9
Cr/weight ratio	-0.255	770	0.000	0.255	10
Waist-to-hip ratio	0.254	773	0.000	0.254	11
HDL-C	-0.25	907	0.000	0.25	12
HbA1c	0.246	495	0.000	0.246	13
Average heart rate	0.233	860	0.000	0.233	14
Hip circumference	0.225	774	0.000	0.225	15
UA	0.22	972	0.000	0.22	16
Kidney meridian	0.208	860	0.000	0.208	17
FPSA	-0.194	444	0.000	0.194	18
WBC	0.191	978	0.000	0.191	19
CRP	0.187	261	0.002	0.187	20
Absolute value of lymphocytes	0.186	981	0.000	0.186	21
Exercise type	-0.185	711	0.000	0.185	22
Intraocular pressure of the left eye	0.185	446	0.000	0.185	23
Glucose	0.183	911	0.000	0.183	24
Liver meridian	0.178	860	0.000	0.178	25
Hemoglobin	0.177	981	0.000	0.177	26
Gallbladder meridian	-0.177	860	0.000	0.177	27
RBC	0.173	980	0.000	0.173	28
Duration of each exercise session	-0.172	700	0.000	0.172	29
Number of exercise sessions per week	-0.171	701	0.000	0.171	30
Pericardium meridian	-0.17	860	0.000	0.17	31
Redness at the sides and tips	-0.169	924	0.000	0.169	32
Spleen meridian	-0.157	860	0.000	0.157	33
PA	0.152	683	0.000	0.152	34
Absolute value of eosinophils	0.15	981	0.000	0.15	35
T3 triiodothyronine	0.15	764	0.000	0.15	36
HCT	0.148	979	0.000	0.148	37
GLO	0.143	769	0.000	0.143	38
Neu/ALB ratio	0.142	929	0.000	0.142	39
Coating at left side of tongue lab-a1	-0.139	924	0.000	0.139	40
Absolute value of neutrophils	0.135	981	0.000	0.135	41
Left side of tongue lab-a	-0.131	924	0.000	0.131	42
Absolute value of monocytes	0.13	981	0.000	0.13	43
Coating at root of tongue lab-a	-0.129	924	0.000	0.129	44
Coating at full tongue lab-a	-0.129	924	0.000	0.129	45
Systolic blood pressure	0.126	786	0.000	0.126	46
Root of tongue lab-a	-0.124	924	0.000	0.124	47
PCT	0.123	889	0.000	0.123	48
Middle of tongue lab-L	0.123	924	0.000	0.123	49
Left side of tongue lab-L	0.122	924	0.000	0.122	50
Coating at left side of tongue lab-L	0.12	924	0.000	0.12	51
PLT	0.119	981	0.000	0.119	52
Coating at tip of tongue lab-a	-0.119	924	0.000	0.119	53
Coating at full tongue lab-l	0.119	924	0.000	0.119	54
TP	0.118	929	0.000	0.118	55
Coating at right side of tongue lab-a	-0.116	924	0.000	0.116	56
Right side of tongue lab-a	-0.115	924	0.000	0.115	57
Tip of tongue lab-l	0.115	924	0.000	0.115	58
Coating at tip of tongue lab-l	-0.114	924	0.000	0.114	59
Coating at middle of tongue lab-l	0.114	924	0.001	0.114	60
Coating at middle of tongue lab-l	0.113	924	0.001	0.113	61
Right side of tongue lab-l	0.109	924	0.001	0.109	62
RDW-SD	-0.107	889	0.001	0.107	63
Coating at middle of tongue lab-a	-0.106	924	0.001	0.106	64
Diastolic blood pressure	0.104	786	0.003	0.104	65
Middle of tongue lab-a	-0.103	924	0.002	0.103	66
Left side of tongue lab-b	-0.103	924	0.002	0.103	67
Coating at right side of tongue lab-l	0.103	924	0.002	0.103	68
Coating at tip of tongue lab-b	-0.102	924	0.002	0.102	69
Absolute value of basophils	0.101	981	0.002	0.101	70
MCHC	0.1	889	0.003	0.1	71
Tip of tongue lab-b	-0.1	924	0.002	0.1	72
Root of tongue lab-b	-0.099	924	0.003	0.099	73
Coating at root of tongue lab-l	0.099	924	0.003	0.099	74
ALP	0.098	798	0.006	0.098	75
Root of tongue lab-l	0.097	924	0.003	0.097	76
Right side of tongue lab-b	-0.096	924	0.004	0.096	77
Coating at left side of tongue lab-b	-0.096	924	0.004	0.096	78
Coating at full tongue lab-b	-0.096	924	0.003	0.096	79
Coating at right side of tongue lab-b	-0.094	924	0.004	0.094	80
MCV	-0.09	889	0.007	0.09	81

¹l value: Indicates the lightness of an image, the higher the value, the greater the brightness; a value: Distinguishes the redness of the tongue body and the thickness of the tongue coating; b value: Distinguishes the white/yellow of the tongue coating, positively correlated, a white coating has a lower overall b value, while a yellow coating has a higher overall b value.

ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; γ-GT: Gamma-glutamyl transferase; TG: Triglyceride; BMI: Body mass index; WC: Waist circumference; UA: Uric acid; Cr: Creatinine; HDL-C: high-density lipoprotein cholesterol; LDL-C: low-density lipoprotein cholesterol; HbA1c: Glycated hemoglobin; FPSA: Free prostate-specific antigen; WBC: White blood cell count; CRP: C-reactive protein; RBC: Red blood cell count; PA: Peak expiratory flow; HCT: Hematocrit; GLO: Globulin; Neu: Neutrophil; ALB: Albumin; PCT: Procalcitonin; PLT: Platelets; TP: Total protein; RDW: Red cell distribution width; MCHC: Mean corpuscular hemoglobin concentration; ALP: Alkaline phosphatase; MCV: Mean corpuscular volume.

Table 2 High linear correlation in 16 indicators: Absence of multicollinearity.

Collinearity diagnostics
Indicators	Sig.	Tolerance	VIF	n	Correlation coefficient	Correlation ranking
WC	0.790	0.558	1.792	774	0.331	7
Duration of each exercise session	0.724	0.824	1.213	700	-0.172	29
Absolute value of lymphocytes	0.647	0.785	1.274	981	0.186	21
RBC	0.199	0.702	1.425	980	0.173	28
MCHC	0.114	0.831	1.203	889	0.1	71
PCT	0.476	0.750	1.334	889	0.123	48
ALT	0.265	0.554	1.806	979	0.506	1
AST/ALT	0.004	0.532	1.878	831	-0.495	2
GLO	0.360	0.880	1.137	769	0.143	38
TG	0.182	0.874	1.144	917	0.358	4
Cr/weight ratio	0.052	0.606	1.649	770	-0.255	10
HbA1c	0.008	0.758	1.319	495	0.246	13
FPSA	0.187	0.845	1.183	444	-0.194	18
Coating at left side of tongue la	0.694	0.866	1.154	924	-0.139	40
Average heart rate	0.415	0.809	1.235	860	0.233	14
Liver meridian	0.774	0.806	1.240	860	0.178	25

WC: Waist circumference; RBC: Red blood cell count; MCHC: Mean corpuscular hemoglobin concentration; PCT: Procalcitonin; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; GLO: Globulin; TG: Triglyceride; Cr: Creatinine; HbA1c: Glycated hemoglobin; FPSA: Free prostate-specific antigen.

Feature selection

After data cleaning based on inclusion and exclusion criteria, 711 individuals with complete data were ultimately included in the ML analysis. The dataset contained 156 indicators (Supplementary Table 2). Among these patients, 504 had MAFLD, including 318 males (77.18%), with a median age of 43.00 years (35.00-53.00). The remaining 207 were non-MAFLD patients, including 151 males (72.95%), with a median age of 44.00 years (36.00-56.00).

In the feature selection process, we employed a dual-strategy approach combining LASSO with RFE methodologies. First, we applied the RFE algorithm to systematically evaluate the initial set of 156 features. We progressively refined our selection through iterative elimination experiments, identifying 15 key predictors (Figure 2A). Next, LASSO regression was applied to enhance selection precision, yielding 12 predictive features (Figure 2B). The importance of each feature in LASSO and its impact on the model inference results are shown in Figure 2C and D. Finally, by intersecting the results from both methods, we established a robust set of 10 core features: AST/ALT ratio, TG, high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), UA/Cr ratio, WC, tongue edge and tip redness, Cr/bodyweight ratio, albumin/globulin (A/G) ratio, and greasy tongue coating (Figure 2E). These variables were selected as the final predictors for model development.

Figure 2 Perform feature selection using LASSO and recursive feature elimination. A: Characteristic indicators screening based on recursive feature elimination (RFE); B: Characteristic indicators screening based on LASSO; C: Feature importance values following LASSO selection; D: Contribution of each feature to the model's inference outcomes; E: Characteristics of RFE combined with LASSO. ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; HDL-C: High-density lipoprotein cholesterol; LDL-C: Low-density lipoprotein cholesterol; TG: Triglyceride; BMI: Body mass index; A/G: Albumin/globulin; RDW: Red Cell Distribution Width; AUROC: Area under the receiver operating characteristic; SVM: Support Vector Machine; GLO: Globulin.

Development of predictive models

Furthermore, we validated the stability and generalisation ability of the predictive models in the internal validation sets. Among the four models, the XGBoost model (AUC = 0.82) had the best predictive effect for MAFLD, followed by the RF model (AUC = 0.79), the SVM model (AUC = 0.81), and the LR model (AUC = 0.78; Figure 3). The accuracy, precision, Recall, and F1-score of each model are shown in Table 3. The XGBoost model demonstrated superior performance in all four aspects. In our study, a 5-fold cross-validation approach was implemented to fine-tune and optimize the parameters of the XGBoost model using the training dataset. Overall, as shown in Figure 4, the model demonstrated high discriminative ability, with AUC values approaching 1 and a mean AUC of 0.918, indicating excellent classification performance across different validation folds.

Figure 3 Performance of machine learning models to predict hepatic steatosis. Receiver operating characteristic curves of the four performing machine learning models. RF: Random Forest; LR: Logistic regression; SVM: Support Vector Machine; ROC: Receiver operating characteristic; AUC: Area under the curve.

Figure 4 Cross-validation receiver operating characteristic curve. The receiver operating characteristic curves for XGBoost model evaluated through five-fold cross-validation. The X-axis represents the false positive rate, and the Y-axis represents the true positive rate. Area under the curve is a measure of the model's overall performance, with higher values indicating better discrimination.

Table 3 Performance comparison of each model.

Models	Accuracy	Precision	Recall	F1-score
Random forest	0.83 (0.7663-0.8804)	0.82 (0.7518-0.8803)	0.83 (0.7717-0.8804)	0.81 (0.7439-0.8692)
XGBoost	0.84 (0.7989-0.9076)	0.82 (0.7987-0.9119)	0.82 (0.7820-0.8533)	0.84 (0.7741-0.8450)
Logistic regression	0.78 (0.7174-0.8370)	0.78 (0.7233-0.8379)	0.78 (0.7120-0.8370)	0.78 (0.7143-0.8340)
Support vector machine	0.78 (0.7227-0.8424)	0.77 (0.7042-0.8372)	0.78 (0.7174-0.8424)	0.78 (0.7100-0.8352)

Feature importance

As shown in Figure 5, a SHAP summary dot plot was utilized to interpret the predictions generated by the XGBoost model. The AST/ALT ratio exhibited the highest impact, followed by LDL-C, A/G ratio, Cr/BW ratio, and TGs. In addition to conventional clinical indicators, TCM features such as Greasy tongue coating and Tongue edge and tip redness also ranked among the top contributors, highlighting the added value of multimodal feature integration in predictive modeling.

Figure 5 SHAP summary dot plot. The probability of metabolic-associated fatty liver disease development increases with the SHAP value of a feature. A dot is made for SHAP value in the model for each single patient, so each patient has one dot on the line for each feature. ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; HDL-C: High-density lipoprotein cholesterol; LDL-C: Low-density lipoprotein cholesterol; TG: Triglyceride; A/G: Albumin/globulin.

DISCUSSION

The global prevalence of MAFLD is increasing annually, primarily due to obesity[22]. Moreover, multiple studies have consistently reported a greater risk of significant hepatic fibrosis in patients with T2DM[23,24]. Screening for MAFLD in all patients with metabolic abnormalities has proven cost-effective[25]. Therefore, the primary purpose of screening for hepatic steatosis in high-risk populations is to implement early intervention measures to prevent cirrhosis development and liver-related all-cause mortality[26,27]. Our study found that 70.8% of individuals with metabolic abnormalities had hepatic steatosis by FibroScan screening. Interestingly, after intergroup comparisons, 81 correlated factors, including laboratory examinations, anthropometric metrics, and TCM constitution-related indicators, were associated with hepatic steatosis in this high-risk population. Is there a direct correlation between these indicators and hepatic fat deposition?

We utilized LASSO and RFE to identify characteristic predictors in this prospective cohort study. After further data cleaning, 711 participants who underwent health check-ups were selected. Among them, 207 (29.2%) were non-MAFLD individuals, and 504 (70.8%) were MAFLD patients. The Hughes phenomenon in high-dimensional spaces necessitates rigorous dimensionality reduction protocols, where feature selection operates under Probably Approximately Correct learning theory to maintain optimal predictive performance while constraining the Vapnik-Chervonenkis dimension[28]. The notion that no single “best method” exists for all problem settings has gained widespread acceptance due to various comparative studies[29,30]. Combining different feature selection methods (hybrid methods) can increase classification accuracy and improve the stability and reproducibility of the results[31]. Our study used LASSO regularization (an embedded method) and RFE (a wrapper method) to screen for feature selection. This rigorous feature selection process combines the strengths of various techniques, ensuring that the selected features have high reliability and robustness. The feature importance ranking indicated that AST/ALT, TG, HDL-C, LDL-C, UA/Cr, WC, tongue edge and tip redness, Cr/BW ratio, A/G, and greasy tongue coating were the top 10 crucial factors. Notably, the correlation analysis and collinearity diagnosis of the original dataset identified 16 indicators strongly associated with MAFLD, including WC, exercise type, exercise duration, absolute lymphocyte count, red blood cell count, mean corpuscular haemoglobin concentration, plateletcrit, ALT, AST/ALT ratio, globulin (GLO), TG, Cr/weight ratio, tongue edge and tip redness, coating at the left side of the tongue (Lab-a), average heart rate, and the liver meridian. In the end, the indicators AST/ALT, TG, WC, and tongue edge and tip redness strongly corresponded with the features identified through RFE combined with LASSO. This consistency not only underscores the clinical rationality of the selection process but also indicates that the chosen features can accurately represent the key risk factors associated with MAFLD.

Using these characteristic predictors, we constructed and validated four ML models for screening for hepatic steatosis in patients at high metabolic risk. The final model we developed, XGBoost, demonstrated superior ability, with an AUC of 0.82, an accuracy of 0.84, a recall of 0.82, a precision of 0.82, and F1-score of 0.84. In clinical practice, although a high AUC value and accuracy indicate that the model maintains high sensitivity and specificity across different thresholds and demonstrates good predictive performance in the overall sample, reliance on these two metrics alone may not be sufficient for a comprehensive assessment of the model's clinical value. Given that MAFLD is typically asymptomatic, our study prioritized the hepatic steatosis in patients at high metabolic risk, rendering recall a particularly crucial metric. Recall, defined as the proportion of true positives among all confirmed cases, measures the model's ability to accurately identify actual hepatic steatosis cases. In this study, the XGBoost model achieved a recall of 0.82, demonstrating its strong performance in detecting hepatic steatosis accurately. Furthermore, a high F1-score, indicates that the model has achieved a good balance between recall and precision, effectively identifying patients with hepatic steatosis while minimizing false positives, thereby providing a more comprehensive reflection of the model's performance in terms of sensitivity and specificity. The cross-validation results revealed an overall mean AUC of 0.918 with a standard deviation of 0.024 across all folds, indicating high consistency in the model's performance.

As health awareness continues to rise and physical examinations become increasingly widespread, a growing number of diseases are now being detected at an early stage through routine check-ups[31]. The reclassification of MAFLD underscores the heterogeneity of hepatic steatosis and its intricate metabolic associations. To date, much research has focused on diagnosing MAFLD[32,33] while neglecting early diagnosis and prediction in high-risk populations. Moreover, most studies have relied on retrospective data or abdominal ultrasound, and the selection of study subjects often lacks specificity[34,35]. This limitation may result in findings that fail to accurately reflect the true characteristics and disease progression patterns of MAFLD across diverse populations, thereby impeding the effectiveness of early prediction and intervention strategies. In this study, we aimed to screen for hepatic steatosis in high-metabolic-risk populations on a large scale. Our approach addresses the limitations of current screening methods, such as ultrasound and MRI-PDFF, in terms of universality and practicality while being more cost-effective regarding medical resources. We evaluated four algorithms (RF, XGBoost, LR, and SVM) to assess the presence of MAFLD. The XGBoost algorithm outperformed the comparative models, achieving an area under the ROC curve of 0.82, indicating robust discriminatory capacity. As a result, considering its optimal balance of these metrics, the XGBoost model proved to be the most effective tool in our analysis for detecting hepatic steatosis.

To enhance the interpretability of our XGBoost model, we employed SHAP analysis. The analysis of feature importance evaluates the impact of each variable on the model's predictive power, highlighting the key determinants that drive the outcomes. The SHAP analysis pinpointed crucial predictors such as the AST/ALT ratio and LDL-C, offering in-depth understanding of how each feature impacts the model's predictions. In addition to conventional clinical indicators, TCM features such as Greasy tongue coating and Tongue edge and tip redness also ranked among the top contributors, highlighting the added value of multimodal feature integration in predictive modeling.

We found that AST/ALT, TG, HDL-C, LDL-C, UA/Cr, WC, tongue edge and tip redness ratio, Cr/BW ratio, A/G, and greasy tongue coating are pivotal factors for hepatic steatosis screening in patients at high metabolic risk. Compared with previous studies, the results of this study align well with those of clinical practice and offer significant practical value. The guidelines identify obesity as the most significant driver of MAFLD. Multiple studies have highlighted the close relationship between BW and MAFLD. On the one hand, obesity is a key risk factor for MAFLD, with the risk of developing the disease, increasing approximately 2.5 times for every 5 kg/m² increase in BMI[36]. Additionally, research conducted in the Chinese population indicates that individuals with a BMI of 28 kg/m² or higher (the obesity threshold) have a significantly greater prevalence of MAFLD than those with a BMI below 24 kg/m² (considered normal weight)[37]. The increase in abdominal fat is closely related to insulin resistance and inflammatory responses, promoting fat accumulation in the liver[37]. Research has shown that for every 10 cm increase in WC, the risk of MAFLD increases by approximately 1.5 times[38]. High BMI and WC are both independent risk factors for developing MAFLD, and the predictive value of combining BMI and WC for MAFLD is greater than that of a single indicator[39]. Reducing BMI and WC variabilities in male patients may accelerate MAFLD remission[40]. TG, HDL-C, and LDL-C are components of lipid profiles in blood tests, and increased lipid accumulation in the blood can lead to an increased risk of MAFLD[41]. In ML modeling, these lipid indicators, combined with other metabolic indicators (such as BMI, blood glucose, and liver function markers), can effectively improve the prediction accuracy of MAFLD[42]. The TG/HDL-C ratio has been proven a powerful indicator for predicting the risk of MAFLD[42]. Studies have shown that ALT levels are significantly higher in patients with MAFLD than in non-MAFLD patients[43]. Elevated ALT levels not only indicate the presence of hepatic steatosis but also suggest more severe liver injury and fibrosis when the AST/ALT ratio is increased[38]. Moreover, TG and the AST/ALT ratio are associated with fatty liver disease and have been used in various diagnostic panels[44,45]. Studies have shown a substantial association between elevated levels of UA/Cr and an increased risk of MAFLD in the Chinese adult population[46,47]. Moreover, the ratio of UA to Cr was positively correlated with the risk of moderate-severe MAFLD. It demonstrated the capacity to differentiate moderate-severe MAFLD from mild MAFLD and non-MAFLD[48]. More detailed studies have shown that the UA/Cr ratio might be of greater concern for the risk of MAFLD in nonobese individuals without T2DM[49]. The A/G ratio, which is the ratio of serum albumin to GLO, indirectly reflects the degree of liver damage. Research has suggested that a combination of indicators, including the albumin-to-γ-GT ratio, can enhance the predictive performance of MAFLD, particularly among nondiabetic patients and women[50]. A comprehensive analysis of these 10 factors revealed that they cover different subtypes of MAFLD, such as varying degrees of severity, differences in patients' body sizes, and the presence or absence of diabetes. Most indicators are associated with metabolic syndrome, which suggests that this model is well suited for screening hepatic steatosis in individuals at high metabolic risk.

While these metabolic indicators provide a robust basis for identifying MAFLD, integrating TCM diagnostics could further enhance the precision of this approach. Specifically, the holistic assessment inherent in TCM may offer additional dimensions and provide complementary insights that augment conventional metabolic and biochemical markers. The diagnostic framework of TCM constitution encompasses a constitution questionnaire, tongue diagnosis, facial diagnosis, and palpation[51]. Among these, tongue diagnosis is an extremely crucial and objective component. In MAFLD patients, the color characteristics of the tongue coating, particularly yellow, are significantly correlated with the composition of the oral and gut microbiomes and are more likely to reflect underlying inflammation and metabolic disorders[52]. This study revealed that computer-aided tongue image analysis technology can increase the diagnostic accuracy of MAFLD and that the use of tongue diagnosis as an auxiliary tool for early screening of MAFLD is noninvasive, convenient, and cost effective[53-55]. Regrettably, the only TCM-related indicators identified were tongue edge and tip redness and greasy tongue coating, which may be associated with the “damp-heat” constitution in TCM. However, these indicators can still provide some insights, as the model’s precision remains within an applicable range even after incorporating TCM indicators. This result may be attributed to the relatively small sample size. Additionally, although we used the same acquisition instruments for standardization, different algorithms may still introduce minor discrepancies in the results.

This study has several limitations. First, the prediction model was constructed based on a single-center population, and its applicability to multicenter groups remains to be validated. In the future, we will actively seek multi-center collaborations to assess the model’s generalizability across different populations and settings. Second, we did not subtype MAFLD, as its heterogeneity may have affected the interpretability of this model. Future research should focus on expanding the sample size and extending long-term follow-up across multiple centers. This will increase model precision and allow the development of predictive models for different complications.

CONCLUSION

The ML model can be used to screen for hepatic steatosis in high-metabolic-risk populations. It may also be one of the few noninvasive tools available for rapidly identifying hepatic steatosis in such patients, thereby significantly improving MAFLD check-up efficiency and reducing related medical expenses.