Using machine learning to identify the most important predictors of fatty liver index in healthy young Taiwanese men

Abstract

Nonalcoholic fatty liver disease (NAFLD) is the most common chronic liver disease worldwide. While many factors have been associated with NAFLD, their relative importance in healthy young populations remains unclear. In this study, we enrolled 7,037 healthy young Taiwanese men aged 20–50 years and applied five machine learning (Mach-L) methods to identify the most important predictors of the fatty liver index (FLI). Two models were constructed: Model 1 included all 28 variables, while Model 2 excluded body fat (BF) to unmask the contribution of other factors. In Model 1, BF was the most important predictor (100% relative importance), followed by serum glutamic pyruvic transaminase (SGPT; 34.54%), high-density lipoprotein cholesterol (HDL-C; 16.51%), age (9.91%), uric acid (UA; 7.17%), and fasting plasma glucose (FPG; 6.25%). In Model 2, after removing BF, the most important predictors were SGPT (100%), HDL-C (36.82%), UA (24.90%), C-reactive protein (21.23%), age (10.05%), and FPG (8.17%). All five machine learning methods outperformed traditional multiple linear regression. These findings highlight the central role of adiposity while also revealing the independent contributions of metabolic, inflammatory, and hepatic factors to FLI in young healthy men.

Introduction

Nonalcoholic fatty liver disease (NAFLD) is the most common chronic liver disease worldwide, affecting approximately 30% of the general population¹. NAFLD is also an independent risk factor for type 2 diabetes, cardiovascular diseases, and mortality^2–4. As the incidence of metabolic diseases increases, the diagnosis of NAFLD may become outdated. Therefore, an international expert panel recently proposed a consensus to redefine fatty liver as metabolic associated fatty liver disease⁵. The new criteria no longer exclude patients with alcoholic fatty liver or viral hepatitis but emphasize the metabolic factors involved in the disease. A recent real-world study by Lin et al.⁶, based on National Health and Nutrition Examination Survey III (NHANES III), confirmed that MAFLD patients exhibit more metabolic disorders than NAFLD patients, suggesting that the new criteria may better identify individuals with metabolic disturbances.

In the past, there have been many factors that were identified related to NAFLD, such as obesity, type 2 diabetes, dietary factors, physical inactivity, age and gender, genetic factors and metabolic syndrome^7,8. Of note, most of these studies took NAFLD as a categorical variable, i.e., whether to have NAFLD (0 and 1). The statistical methods used in these studies could only find the risk factors in two groups which is relatively rough compared to fatty liver index (FLI) which transforms the categorical variable into a continuous variable. In this way, more subtle relationships between risk factors and NAFLD could be revealed.

Recent systematic reviews have estimated the global prevalence of NAFLD at approximately 30%, with higher rates observed in individuals with obesity and type 2 diabetes^1,9. Machine learning approaches have been increasingly applied to predict NAFLD risk, with studies demonstrating superior performance compared to traditional regression methods^10–12. For instance, Chen et al.¹⁰ used random forest to identify key predictors of NAFLD in a Chinese cohort, while Yip et al.¹¹ applied gradient boosting machines to develop a screening tool for advanced fibrosis. The FLI, first proposed by Bedogni et al.¹³, has been validated across multiple populations as a reliable non-invasive marker of hepatic steatosis^14,15.

Recently artificial intelligence using machine learning (Mach-L) methods has developed rapidly and is used widely in medical research. Computer algorithms are applied in these methods which could improve the accuracy of prediction through the experiences from analyzing the studied data repeatedly. Some studies focused in this area began to use Mach-L¹⁶.

The five machine learning algorithms selected for this study—Random Forest (RF), stochastic gradient boosting (SGB), multivariate adaptive regression splines (MARS), eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM)—were chosen based on several considerations. First, these methods represent diverse families of machine learning approaches: ensemble tree-based methods (RF, SGB, XGBoost, LightGBM) and spline-based nonlinear regression (MARS). This diversity allows us to compare performance across fundamentally different algorithmic paradigms. Second, these methods have been widely and successfully applied in medical prediction tasks, demonstrating robust performance with mixed data types (continuous and categorical variables)^17–19. Third, they each offer built-in mechanisms for variable importance ranking, which was essential for our primary objective of identifying the most important predictors of FLI. Fourth, these algorithms are capable of capturing complex, non-linear relationships and interactions among variables without requiring prior assumptions about data distribution. Finally, they have demonstrated superior performance in previous NAFLD prediction studies^10–12. While other algorithms such as support vector machines or neural networks could potentially be applied, the selected methods provide a comprehensive and interpretable foundation for our analytical goals.

In the present study, we used five different Mach-L methods to identify which were the most important factors influencing FLI level in healthy young Taiwanese men. Our goals were: (1) To compare the accuracy between traditional multiple linear regression (MLR) and Mach-L. (2) By using FLI, to identify the most relevant factors for NAFLD in Taiwanese men between 20 and 50 years old (included).

Materials and methods

Patient selection

Some of the following contexts were published in our previous study²⁰. Data for this study were sourced from the Taiwan MJ cohort, an ongoing prospective cohort of health examinations conducted by the MJ Health Screening Centers in Taiwan. These examinations cover more than 100 important biological indicators, including anthropometric measurements, blood tests, imaging tests, etc. Each participant completed a self-administered questionnaire to collect information on personal and family medical history, current health status, lifestyle, physical exercise, sleep habits, and dietary habits. The data of our study participants were obtained from MJ clinic. At the time when the participants had the health checkup, general consent forms for future anonymous study were obtained. This database was kept and maintained by the Interpretation Foundation of MJ clinic. All or part of the data used in this research were authorized by and received from MJ Health Research Foundation (Authorization Code: MJHRF2024021A). Any interpretations or conclusions described in this paper do not represent the views of MJ Health Research. The present study is a secondary database analysis and does not involve sample collection, so there is no consent form for this research project. For the data collection procedures, please refer to the technical report annually MJ Health Research Foundation²¹. The present study was approved by the Institutional Review Board of the Tri-Service General Hospital (IRB No.: C202405017) and all methods were performed in accordance with the relevant guidelines and regulations of our institution. As it was a secondary study, only a short review was taken. It is important to note that this study focuses exclusively on men. Given the well-established sex-based differences in body fat distribution, hormonal regulation, and metabolic risk profiles, we chose to analyze men and women separately to avoid confounding and to provide sex-specific insights. The present study focuses on the male cohort; a parallel analysis of the female cohort is currently underway and will be reported separately.

In the beginning, 1,498,312 subjects were enrolled. After excluding subjects who did not fit our criteria, 7,037 men were left for further analysis (Fig. 1). The inclusion criteria were:

Fig. 1.

The flow chart of selecting the study participants.

1. Men between 20 and 50 years old (included).

2. Without any history of significant medical diseases such as stroke, myocardial infarction, or heart failure.

3. Without diagnosis of diabetes.

4. Without taking medications for metabolic syndrome.

To achieve our study objectives, we constructed two distinct predictive models:

Model 1. This model included all available variables listed in Table 1, with body fat (BF) as one of the predictors. This initial model aimed to identify the most important factors for FLI from the full set of clinical, demographic, and lifestyle parameters.

Table 1.

Demographic data of participants.

Numeric variable	Unit	Mean ± SD
Age	Years	36.89 ± 7.73
Body fat (BF)	%	25.3 ± 6.2
Systolic blood pressure (SBP)	mmHg	121.67 ± 14.09
Diastolic blood pressure (DBP)	mmHg	78.89 ± 10.47
Leukocyte (WBC)	*103/µL	6.24 ± 1.59
Hemoglobin (Hb)	*106/µL	15.43 ± 1.08
Platelets (Plt)	*103/µL	233.33 ± 50.4
Fasting plasma glucose (FPG)	mg/dL	100.41 ± 16.14
Total bilirubin (TBIL)	mg/dL	1.04 ± 0.41
Albumin (Alb)	mg/dL	4.52 ± 0.22
Globulin (Glo)	g/dL	3.03 ± 0.32
Alkaline phosphatase (ALP)	IU/L	61.17 ± 17.5
Serum glutamic oxaloacetic transaminase (SGOT)	IU/L	25.62 ± 14.05
Serum glutamic pyruvic transaminase (SGPT)	IU/L	36.37 ± 29.82
Lactate dehydrogenase (LDH)	mg/dL	161.85 ± 35.89
Estimated Glomerular filtration rate (eGFR)	mg/dL	84.84 ± 11.78
Uric acid (UA)	mg/dL	6.62 ± 1.32
High density lipoprotein cholesterol (HDL-C)	mg/dL	52.06 ± 11.56
Low density lipoprotein cholesterol (LDL-C)	mg/dL	125.25 ± 33.48
C reactive protein (CRP)	mg/dL	5.69 ± 15.44
Smoking area	-	7.18 ± 16.23
Chewing betel nut (CBN) area	-	0.55 ± 5.42
Sport area	-	6.55 ± 8.29
Vitamin D (Vit D)	nmol/L	23.49 ± 7.68

Ordinal variables		N (%)
Marriage status (MS)
(1) Unmarried		2362(41.12%)
(2) Married		3382(58.88%)
Income level (IL)**
(1) Below $200,000		413(6.35%)
(2) $200,001–$400,000		631(9.71%)
(3) $400,001–$800,000		2301(35.39%)
(4) $800,001–$1,200,000		1707(26.26%)
(5) $1,200,001–$1,600,000		639(9.83%)
(6) $1,600,001–$2,000,000		325(5.00%)
(7) More than $2,000,000		485(7.46%)
Education level (Edu.)**
(1) Illiterate		1(0.02%)
(2) Elementary school		4(0.06%)
(3) Junior high school		58(0.88%)
(4) High school (vocational)		849(12.86%)
(5) Junior college		736(11.15%)
(6) University		3270(49.55%)
(7) Graduate school or above		1682(25.48%)
Sleeping hours (SH)
(1) 0 ~ 4 h		55(0.81%)
(2) 4 ~ 6 h		1834(26.89%)
(3) 6 ~ 7 h		3479(51.01%)
(4) 7 ~ 8 h		1238(18.15%)
(5) 8 ~ 9 h		174(2.55%)
(6) more than 9 h		40(0.59%)

Dependent variable	Unit	Mean ± SD
Non-alcoholic fatty liver disease (NAFLD) index a	-	33.31 ± 26.74

*By using t-test, there was a significance between unmarried and married participants. ** Indicates that there were significant difference in different education and income level by one way analysis of variance Bonferroni was used as post-Hoc test. The details of the difference between these levels are not shown due to it is complicated.

^a Non-alcoholic fatty liver disease (NAFLD) index.

.

Model 2. This model was developed by removing BF from the predictor set. The rationale for this was to uncover the influence of other clinically relevant variables that might be statistically ‘masked’ by the overwhelming predictive power of BF, as was observed in the preliminary analysis of Model 1 (detailed in the Results section).

After the first analysis including all the variables (Model 1), our result showed that BF was overwhelmingly important (the importance percentage = 100%). This result might ‘mask’ the importance of other variables; therefore, Model 2 was built by removing BF so that we could further detect which parameters were more important.

Measurements of anthropometry and biochemistry

Part of the following content was published by our team and it could be referred to²². On the day of examination, a senior nurse recorded the people’s personal history including smoking, alcohol drinking, education level, betel nut use. Body weight (kilogram) was examined by electronic body weight meter. Systolic blood pressure (SBP) and diastolic blood pressure (DBP) were measured using an electric sphygmomanometer.

Blood samples were obtained after a 10-hour fasting period. Within one hour of collection, plasma was separated from the blood and stored at −30 degrees Celsius until lipid profile analysis. Total cholesterol and triglyceride (TG) levels were measured using a dry, multilayer analytical slide method, employing the Fuji Dri-Chem 3000 analyzer (Fuji Photo Film, Tokyo, Japan). Serum high-density lipoprotein cholesterol (HDL-C) and low-density lipoprotein cholesterol (LDL-C) concentrations were determined through an enzymatic cholesterol assay following dextran sulfate precipitation.

Traditional statistics

Data are represented as means ± standard deviations. t-test was used in testing the difference between married and unmarried participants. At the same time, the education and income levels were ordinal variables and analysis of variance (ANOVA) was applied. Pearson’s correlation was used to analyze the relationships between all the continuous variables and FLI. MLR was performed here and was taken as a benchmark for comparison to Mach-L. All statistical tests were two-sided, and p < 0.05 was considered statistically significant. Statistical analysis was performed using SPSS 10.0 for Windows (SPSS, Chicago, IL, USA).

Description of the study data set

In this study, Table 1 provides the mean ± standard deviation for all clinical variables. The smoking area was the multiple of the duration, frequency of smoking, and number of cigarettes. The sport area was the multiple of duration, frequency, and type of exercise. The sleep time was an ordinal variable as shown in Table 1. Finally, the equation of FLI = e^y^/(1 + e^y^) × 100, where y = 0.953 × ln(triglycerides, mg/dL) + 0.139 × (BMI kg/m²) + 0.718 × ln(GGT, U/L) + 0.053 × (waist circumference, cm) − 15.745²³.

Proposed machine learning scheme

This study proposes a scheme based on five different Mach-L methods, namely, RF, SGB, MARS, XGBoost, and LightGBM to identify the most important factors influencing FLI. These Mach-L methods have been widely applied in various healthcare applications and do not rely on prior assumptions regarding data distribution^14,15. This part of method has been published by our group¹⁶.

RF is an ensemble learning decision trees algorithm that combines bootstrap resampling and bagging²⁴. Its principle involves randomly generating multiple different and unpruned CART decision trees, where the decrease in Gini impurity is used as the splitting criterion. All the generated trees are then combined to form a forest. The final model is obtained by averaging or voting among all the trees in the forest to generate output probabilities and ensure robustness²⁵.

The next method, SGB, is a tree-based gradient boosting learning algorithm that combines bagging and boosting techniques to minimize the loss function and address the over-fitting problem of traditional decision trees²⁶. In SGB, it sequentially generates many stochastic weak learners (trees) through multiple iterations. Each tree focuses on correcting or explaining errors from the tree of the previous iteration, with the residual of the previous iteration tree serving as the input for the newly generated tree. This iterative process continues until a stopping criterion is reached, such as the maximum number of iterations or the convergence condition. Ultimately, the cumulative results of many trees are used to determine the final robust model.

The third one, MARS is a nonparametric and nonlinear statistical method in which several linear segments with different gradients are used to automatically examine the nonlinearity and dependency between multidimensional input and output variables, and then generate the final optimum nonlinear prediction model²⁶.

The fourth method, XGBoost, is a gradient boosting technology based on the optimized extension of SGB²⁷. The underlying principle of XGBoost involves training many weak models sequentially and assembling them using the gradient boosting method to achieve improved prediction performance. To accelerate the model construction convergence process, XGBoost employs the Taylor binomial expansion to approximate the objective function and utilizes arbitrary differentiable loss functions. Additionally, XGBoost applies a regularized boosting technique to penalize the complexity of the model, thereby correcting overfitting and increasing model accuracy²⁷.

Finally, LightGBM is a decision tree-based distributed gradient boosting framework that uses advanced histograms. In each iteration, it learns the approximate value of the decision tree residuals based on one-side sampling and negative gradient fitting²⁸.

Figure 2 presents the flowchart of the proposed prediction and important variable identification scheme that combines the five Mach-L methods. Initially, patient data were collected and used to prepare the dataset through the proposed method. Subsequently, the dataset was randomly divided into an 80% training dataset for model building and a 20% testing dataset for model evaluation. During the training process, each Mach-L method was fine-tuned with specific hyperparameters to construct well-performing models. To achieve this, a 10-fold cross-validation (CV) technique was employed for hyperparameter tuning. The training dataset was further partitioned into a training subset with a different set of hyperparameters and a validation subset for model validation. A grid search explored all possible combinations of hyperparameters. The model with the lowest root mean square error for the validation subset was deemed the best model for each Mach-L method. Consequently, the best-tuned RF, SGB, MARS, XGBoost, and LightGBM models were generated, and the corresponding variable importance ranking information was obtained.

Fig. 2.

Proposed machine learning prediction scheme.

For the testing process, the testing dataset was utilized to assess the predictive performance of the best RF, SGB, and XGBoost models. Since the target variable in this study is a numerical variable, model performance comparison utilized metrics such as the symmetric mean absolute percentage error (SMAPE), relative absolute error (RAE), root relative squared error (RRSE), and root mean squared error (RMSE), which are presented in Tables 2 and 3. SMAPE is a useful, balanced way to verify the mean percentage of error for predicted results. The basal MAPE formula places only the absolute value of the actual value into the denominator, which gives the restriction that no instance may be the actual value is zero or very small. As an alternative, SMAPE was used. The RAE is calculated as the sum of the absolute differences between the predicted and actual values, divided by the sum of the absolute differences between the actual values and the mean of the actual values. This metric provides an intuitive way to assess the accuracy of a model in relation to a simple baseline, making it useful for comparing different prediction methods. RMSE is calculated as the square root of the average of the squared differences between the predicted and actual values. RMSE penalizes larger errors more heavily, making it sensitive to outliers and providing a clear indication of the model’s accuracy. RRSE is determined by dividing the sum of the squared differences between the predicted and actual values by the sum of the squared differences between the actual values and the mean of the actual values, and then taking the square root of the result. Lower values of SMAPE, RAE, RRSE, and RMSE indicate better prediction, as they signify that the model’s predicted values are closer to the actual values.

Table 2.

The average performance of the multivariate adaptive regression splines and multiple linear regression methods in model without body fat (Model 2).

Methods	MAPE	SMAPE	RAE	RRSE	RMSE
MLR	0.952	0.567	0.617	0.675	18.039
RF	0.837	0.484	0.581	0.637	17.026
SGB	0.795	0.523	0.566	0.629	16.810
MARS	0.899	0.546	0.580	0.640	17.120
XGBoost	0.804	0.527	0.567	0.629	16.812
LightGBM	0.719	0.462	0.559	0.634	16.945

Data showed as mean; MLR multiple linear regression, RF Random Forest, SGB stochastic gradient boosting, MARS multivariate adaptive regression splines, XGBoost eXtreme gradient boosting, LightGBM Light Gradient Boosting Machine, MAPE Mean Absolute Percentage Error; SMAPE symmetric mean absolute percentage error, RAE Relative Absolute Error, RRSE Root Relative Squared Error, RMSE Root Mean Squared Error.

Table 3.

The percentage of importance of variables in Model 1 (with body fat).

Rank	Variable	RF	SGB	MARS	XGBoost	LightGBM	AVG
1	BF	100	100	100	100	100	100
2	SGPT	38.41	28.51	46.74	38.58	20.45	34.54
3	HDL-C	14.6	11.91	32.26	14.68	9.12	16.51
4	Age	6.64	6.35	24.81	5.45	6.32	9.91
5	UA	6.17	5.1	15.15	5.58	3.86	7.17
6	FPG	4.94	3.91	15.15	3.72	3.53	6.25
7	WBC	3.87	1.83	7.75	1.7	2.45	3.52
8	DBP	4.38	3.46	0	4.39	3.04	3.05
9	SBP	3.44	1.65	4.28	1.97	2.06	2.68
10	CRP	6.03	1.1	0	4.24	1.08	2.49
11	LDL-C	3.96	1.47	0	1.6	2.48	1.9
12	SGOT	5.54	0.79	0	0.96	1	1.66
13	TBIL	3.11	1.32	0	1.52	1.9	1.57
14	Smoking area	1.24	1.54	0	1.23	1.39	1.08
15	Plt	2.41	0.55	0	1.34	1.01	1.06
16	LDH	2.22	0.44	0	0.78	1.18	0.92
17	Hb	2.04	0.41	0	0.6	0.68	0.75
18	eGFR	2.02	0.37	0	0.75	0.61	0.75
19	ALP	2.03	0.3	0	0.56	0.8	0.74
20	Vit D	1.8	0.32	0	0.61	0.73	0.69
21	Alb	1.47	0.43	0	0.52	0.62	0.61
22	Glo	1.35	0.11	0	0.25	0.42	0.43
23	Sport area	1.13	0.07	0	0.14	0.45	0.36
24	IL	0.83	0.25	0	0.26	0.31	0.33
25	Edu.	0.57	0.13	0	0.31	0.16	0.23
26	SH	0.35	0.06	0	0.3	0.22	0.19
27	CBN area	0	0.33	0	0.25	0.19	0.15
28	MS	0.2	0	0	0	0	0.04

MS Marriage status, IL Income level, Edu. Education level, BF Body fat, SBP Systolic blood pressure, DBP Diastolic blood pressure, WBC Leukocyte, Hb Hemoglobin, Plt Platelets, FPG Fasting plasma glucose, TBIL Total bilirubin, Alb Albumin, Glo Globulin, ALP Alkaline phosphatase, SGOT Serum glutamic oxaloacetic transaminase, SGPT Serum glutamic pyruvic transaminase, LDH Lactate dehydrogenase, eGFR estimated Glomerular filtration rate, UA Uric acid, HDL-C High density lipoprotein cholesterol, LDL-C Low density lipoprotein cholesterol, CRP C reactive protein, CBN area Chewing betel nut area, SH Sleeping hours, Vit D Vitamin D.

To ensure a robust comparison, the training and testing processes mentioned above were repeated randomly for 10 iterations. The averaged metrics of the RF, SGB, and XGBoost models were then utilized to compare the model performance against the benchmark MLR model, which used the same training and testing datasets as the Mach-L methods. A Mach-L model with an average metric lower than that of MLR was considered a convincing model.

Since all the Mach-L methods used in this study can produce importance rankings of each predictor variable, we assigned priority to each variable based on its ranking in each model, with rank 1 being the most critical risk factor and rank 21 being the least important. However, different Mach-L methods may yield distinct variable importance rankings due to their unique modeling characteristics. To enhance the stability and integrity of re-ranking the importance of risk factors, we integrated the variable importance rankings from the convincing Mach-L models.

In the final stage of the proposed scheme, we summarized and discussed the significant findings based on the results from the convincing Mach-L models. This process allowed us to identify important variables that contribute significantly to predicting FLI.

Machine learning hyperparameters

To ensure optimal model performance, hyperparameter tuning was conducted for each machine learning algorithm using a grid search approach with 10-fold cross-validation on the training dataset. The final hyperparameter sets selected based on the lowest RMSE were as follows:

RF

• Number of trees (ntree): 500.

• Number of variables randomly sampled as candidates at each split (mtry): 5.

• Minimum size of terminal nodes (nodesize): 5.

• Maximum depth of trees (maxdepth): 10.

SGB

• Number of trees (n.trees): 300.

• Interaction depth (interaction.depth): 4.

• Shrinkage (learning rate): 0.01.

• Minimum observations per node (n.minobsinnode): 10.

MARS

• Maximum degree of interaction (degree): 2.

• Penalty per knot (penalty): 2.

• Number of terms retained after pruning (nprune): 20.

XGBoost

• Number of rounds (nrounds): 200.

• Maximum tree depth (max_depth): 4.

• Learning rate (eta): 0.05.

• Minimum child weight (min_child_weight): 3.

• Subsample ratio (subsample): 0.8.

• Column subsample ratio (colsample_bytree): 0.8.

LightGBM

• Number of leaves (num_leaves): 31.

• Learning rate (learning_rate): 0.05.

• Feature fraction (feature_fraction): 0.8.

• Bagging fraction (bagging_fraction): 0.8.

• Bagging frequency (bagging_freq): 5.

• Minimum data in leaf (min_data_in_leaf): 20.

These hyperparameters were selected after extensive grid search and cross-validation to balance model complexity with predictive accuracy while avoiding overfitting.

In this study, all methods were performed using R software version 4.0.5 and RStudio version 1.1.453 with the required packages installed^29,30. The implementations of RF, SGB, and XGBoost were the “randomForest” R package version 4.6–14³¹, “gbm” R package version 2.1.8³², and “XGBoost” R package version 1.5.0.2, respectively³³. In addition, to estimate the best hyperparameter set for the developed effective RF, SGB, and XGBoost methods, the “caret” R package version 6.0–90 was used³⁴. MARS was implemented using the “earth” R package. The MLR was implemented using the “stats” R package version 4.0.5, and the default setting was used to construct the models.

Results

Demographic

A total of 7,037 healthy young Taiwanese men aged between 20 and 50 years were enrolled in this study. Their demographic, biological, and lifestyle characteristics are presented in Table 1. The mean age of the cohort was 36.89 ± 7.73 years, with a mean body fat percentage of 25.3% ± 6.2% and a mean FLI of 33.31 ± 26.74, indicating a wide range of fatty liver risk within this healthy population. Regarding lifestyle factors, the mean smoking area was 7.18 ± 16.23, and the mean sport area was 6.55 ± 8.29, reflecting considerable variability in health behaviors among participants.

Significant differences were observed between married and unmarried participants, as well as across different education and income levels (Table 1). However, these demographic details, while statistically significant, are less central to the primary aims of this study and have been summarized here without exhaustive elaboration.

Results of simple correlation

Table 4 displays the simple Pearson correlation coefficients between all continuous independent variables and FLI. Several variables demonstrated significant positive correlations with FLI, including age (r = 0.166, p < 0.001), BF (r = 0.745, p < 0.001), SBP (r = 0.348, p < 0.001), DBP (r = 0.348, p < 0.001), white blood cell count (WBC; r = 0.345, p < 0.001), fasting plasma glucose (FPG; r = 0.278, p < 0.001), serum glutamic pyruvic transaminase (SGPT; r = 0.456, p < 0.001), uric acid (UA; r = 0.390, p < 0.001), and LDL-C (r = 0.266, p < 0.001). These findings suggest that adiposity, glycemic dysregulation, liver enzymes, and inflammatory markers are all positively associated with fatty liver risk at the univariate level.

Table 4.

Results of simple correlation between all continuous independent variables and fatty liver index.

	Age	BF	SBP	DBP	WBC	Hb	Plt
FLI	0.166***	0.745***	0.348***	0.348***	0.345***	0.203***	0.190***

	FPG	TBIL	Alb	Glo	ALP	SGOT	SGPT
FLI	0.278***	− 0.214***	− 0.003	0.160***	0.158***	0.284***	0.456***

	LDH	eGFR	UA	HDL-C	LDL-C	CRP	Smoking area
FLI	0.192***	− 0.055***	0.390***	− 0.432***	0.266***	0.157***	0.158***

	CBN area	Sport area	SH	Vit D
FLI	0.112***	− 0.171***	− 0.057***	− 0.061***

FLI Fatty liver index, BF Body fat, SBP Systolic blood pressure, DBP Diastolic blood pressure, WBC Leukocyte, Hb Hemoglobin, Plt Platelets, FPG Fasting plasma glucose,TBIL Total bilirubin, Alb Albumin, Glo Globulin, ALP Alkaline Phosphatase, SGOT Serum glutamic oxaloacetic transaminase, SGPT Serum glutamic pyruvic transaminase, LDH Lactate dehydrogenase, eGFR estimated Glomerular filtration rate, UA Uric acid, HDL-C High density lipoprotein cholesterol, LDL-C Low density lipoprotein cholesterol, CRP C reactive protein, CBN area Chewing betel nut area, SH Sleeping hours, Vit D Vitamin D. ***: p < 0.001.

Conversely, several variables exhibited significant negative correlations with FLI, most notably HDL-C (r = −0.432, p < 0.001), total bilirubin (TBIL; r = −0.214, p < 0.001), estimated glomerular filtration rate (eGFR; r = −0.055, p < 0.001), sport area (r = −0.171, p < 0.001), and vitamin D (r = −0.061, p < 0.001). The strong inverse relationship between HDL-C and FLI is particularly noteworthy, underscoring the potential protective role of HDL-C against hepatic fat accumulation.

As shown in Table 4, serum albumin exhibited a negligible and statistically non-significant correlation with FLI (r = −0.003, p > 0.05), indicating no meaningful association in this cohort. This finding is consistent with albumin’s role as a non-specific marker of nutritional status rather than a direct correlate of hepatic steatosis.

Comparison of accuracy between MLR and Mach-L

As described in the Methods section, two analytical models were constructed: Model 1, which included all 28 variables including BF, and Model 2, which was built after removing BF to unmask the relative importance of other factors. The predictive performance of the five Mach-L methods— RF, SGB, MARS, XGBoost, and LightGBM—was compared against traditional MLR as a benchmark. The evaluation metrics used were SMAPE, RAE, RRSE, and RMSE, with lower values indicating superior predictive accuracy.

The results for Model 1 are presented in Table 5. A clear and consistent finding was that all five Mach-L methods outperformed the traditional MLR model. For example, the RMSE for MLR was 14.452, while the RMSE values for the Mach-L methods ranged from 13.122 (SGB) to 13.565 (RF), representing a meaningful improvement in prediction accuracy. Similarly, the SMAPE for MLR was 0.495, whereas LightGBM achieved the lowest SMAPE of 0.378, followed closely by RF at 0.383. These results demonstrate that machine learning approaches capture complex, non-linear relationships in the data that a linear model like MLR cannot adequately model.

Table 5.

The average performance of the multivariate adaptive regression splines and multiple linear regression methods in model with body fat (Model 1).

Methods	MAPE	SMAPE	RAE	RRSE	RMSE
MLR	0.838	0.495	0.507	0.548	14.452
RF	0.532	0.383	0.456	0.514	13.565
SGB	0.544	0.430	0.444	0.497	13.122
MARS	0.668	0.471	0.467	0.507	13.367
XGBoost	0.569	0.442	0.451	0.505	13.320
LightGBM	0.498	0.378	0.439	0.504	13.298

Data showed as mean; MLR multiple linear regression, RF Random Forest, SGB stochastic gradient boosting, MARS multivariate adaptive regression splines, XGBoost eXtreme gradient boosting, LightGBM Light Gradient Boosting Machine, MAPE Mean Absolute Percentage Error, SMAPE symmetric mean absolute percentage error, RAE Relative Absolute Error, RRSE Root Relative Squared Error, RMSE Root Mean Squared Error.

When comparing performance among the five machine learning models in Model 1, some nuanced differences were observed. LightGBM and RF demonstrated the strongest predictive capability, achieving the lowest SMAPE (0.378 and 0.383, respectively) and MAPE (0.498 and 0.532, respectively). SGB and XGBoost also performed very well, with nearly identical error metrics across all four evaluation measures. MARS, while still superior to MLR, was generally the least performant among the five machine learning methods.

The results for Model 2 (without BF) are presented in Table 2. Consistent with Model 1, all Mach-L methods demonstrated superior predictive accuracy compared to MLR. The MLR model’s RMSE of 18.039 was substantially higher than those of the Mach-L methods, which ranged from 16.810 (SGB) to 17.120 (MARS). LightGBM again showed the strongest performance, achieving the lowest SMAPE (0.462) and MAPE (0.719), followed closely by RF, SGB, and XGBoost.

Interpretation of model performance

The consistent superiority of ensemble tree-based methods (RF, SGB, XGBoost, LightGBM) over both MLR and MARS warrants explanation. These algorithms are specifically designed to capture complex, non-linear relationships and high-order interactions among variables without requiring prior specification. In the context of NAFLD, where metabolic, inflammatory, and lifestyle factors interact in intricate ways, this flexibility is a significant advantage. For instance, the relationship between uric acid and FLI may be modified by age or renal function—interactions that tree-based methods can automatically detect and model.

LightGBM’s consistently strong performance across both models highlights its efficiency and robustness. As a gradient boosting framework that uses leaf-wise tree growth, LightGBM can capture finer-grained patterns in the data compared to level-wise growth methods. Its advanced histogram-based algorithms also contribute to stable and accurate predictions.

MARS, while more flexible than MLR due to its piecewise linear splines, was generally the least performant among the five machine learning methods. This suggests that the relationships between predictors and FLI are not merely non-linear but involve complex interactions that are better captured by ensemble methods. MARS can model non-linearity through hinge functions but does not naturally capture the type of high-order interactions that tree-based ensembles excel at.

These findings reinforce the value of boosting and bagging techniques for capturing the intricate interactions among metabolic, inflammatory, and clinical variables that contribute to fatty liver disease. The consistent outperformance of all five machine learning methods compared to traditional linear regression underscores the multifactorial and non-linear nature of NAFLD pathophysiology.

Main results of Mach-L

Model 1 (with BF)

Table 3 presents the percentage importance of each variable derived from the five Mach-L methods in Model 1, along with their averaged values. As anticipated, BF emerged as the overwhelmingly most important predictor, with a relative importance of 100% across all five algorithms. This finding is consistent with the strong simple correlation observed in Table 4 (r = 0.745, p < 0.001) and aligns with the well-established central role of adiposity in the pathogenesis of NAFLD.

After BF, the next five most important factors, in descending order of average importance, were: SGPT (34.54%), HDL-C (; 16.51%), age (9.91%), UA (7.17%), and FPG (6.25%). These results indicate that, even after accounting for the dominant effect of BF, liver enzyme levels, lipid metabolism, chronological age, and glycemic control remain critical determinants of FLI in this young male cohort. The prominence of SGPT—a liver-specific enzyme—underscores its value as a marker of hepatocellular injury, while the strong showing of HDL-C reinforces its protective role against hepatic steatosis. A graphical representation of these importance rankings is provided in Fig. 3.

Fig. 3.

The figure illustration by the order of percentage of importance of the variables for Model 1 (with body fat). MS: Marriage status; IL: Income level; Edu.: Education level; BF: Body fat; SBP: Systolic blood pressure; DBP: Diastolic blood pressure; WBC: Leukocyte; Hb: Hemoglobin; Plt: Platelets; FPG: Fasting plasma glucose; TBIL: Total bilirubin; Alb: Albumin; Glo: Globulin; ALP: Alkaline phosphatase; SGOT: Serum glutamic oxaloacetic transaminase; SGPT: Serum glutamic pyruvic transaminase; LDH: Lactate dehydrogenase; eGFR: estimated Glomerular filtration rate; UA: Uric acid; HDL-C: High density lipoprotein cholesterol; LDL-C: Low density lipoprotein cholesterol; CRP: C reactive protein; CBN area: Chewing betel nut area; SH: Sleeping hours; Vit D: Vitamin D.

Model 2 (without BF)

To avoid the possibility that the overwhelming importance of BF in Model 1 might mask the contribution of other clinically relevant variables, we constructed Model 2 with BF removed. The results are shown in Table 6 and illustrated in Fig. 4.

Table 6.

The percentage of importance of variables in Model 2 (without body fat).

Rank	Variable	RF	SGB	MARS	XGBoost	LightGBM	AVG
1	SGPT	100	100	100	100	100	100
2	HDL-C	37.07	31.48	60.08	26.94	28.52	36.82
3	UA	23.09	20.15	45.85	18.23	17.19	24.9
4	CRP	23.68	19.49	25.41	18.88	18.69	21.23
5	SBP	13.23	10.86	37.48	11.04	11.28	16.78
6	WBC	13.21	11.65	23.57	9.44	10.61	13.7
7	Age	8.4	5.83	23.57	5.73	6.72	10.05
8	FPG	9.88	10.37	5.97	6.53	8.08	8.17
9	LDL-C	10.77	6.79	0	5.25	8.08	6.18
10	DBP	9.65	6.01	0	3.63	5.44	4.95
11	SGOT	12.95	2.43	0	2.5	2.25	4.03
12	TBIL	6.12	3.12	0	2.23	3.79	3.05
13	Plt	5.77	1.9	0	1.49	3.41	2.51
14	LDH	5.02	1.93	0	1.39	3.37	2.34
15	Hb	4.95	1.55	0	1.06	1.85	1.88
16	Smoking area	1.85	2.56	0	2.15	2.36	1.78
17	eGFR	5.03	0.65	0	0.65	2.45	1.76
18	ALP	4.28	1.09	0	0.28	1.51	1.43
19	Alb	2.97	1.02	0	0.7	1.85	1.31
20	Glo	2.76	0.6	0	0.37	0.91	0.93
21	CBN area	0.06	2.1	0	1.59	0.91	0.93
22	Sport area	2.58	0.24	0	0.11	1.46	0.88
23	IL	1.39	0.61	0	0.08	0.63	0.54
24	Edu.	1.02	0	0	0.09	0.23	0.27
25	MS	0	0.46	0	0.17	0.65	0.26
26	SH	0.19	0	0	0	0	0.04

MS Marriage status,IL Income level, Edu. Education level, SBP Systolic blood pressure, DBP Diastolic blood pressure, WBC Leukocyte, Hb Hemoglobin, Plt Platelets, FPG Fasting plasma glucose, TBIL Total bilirubin, Alb Albumin, Glo Globulin, ALP Alkaline phosphatase, SGOT Serum glutamic oxaloacetic transaminase, SGPT Serum glutamic pyruvic transaminase, LDH Lactate dehydrogenase, eGFR estimated Glomerular filtration rate, UA Uric acid, HDL-C High density lipoprotein cholesterol,LDL-C Low density lipoprotein cholesterol, CRP C reactive protein, CBN area Chewing betel nut area, SH Sleeping hours.

Fig. 4.

The figure illustration by the order of percentage of importance of the variables for Model 2 (without body fat). MS: Marriage status; IL: Income level; Edu.: Education level; BF: Body fat; SBP: Systolic blood pressure; DBP: Diastolic blood pressure; WBC: Leukocyte; Hb: Hemoglobin; Plt: Platelets; FPG: Fasting plasma glucose; TBIL: Total bilirubin; Alb: Albumin; Glo: Globulin; ALP: Alkaline phosphatase; SGOT: Serum glutamic oxaloacetic transaminase; SGPT: Serum glutamic pyruvic transaminase; LDH: Lactate dehydrogenase; eGFR: estimated Glomerular filtration rate; UA: Uric acid; HDL-C: High density lipoprotein cholesterol; LDL-C: Low density lipoprotein cholesterol; CRP: C reactive protein; CBN area: Chewing betel nut area; SH: Sleeping hours; Vit D: Vitamin D.

With BF removed, SGPT emerged as the most important factor, achieving 100% relative importance across all five Mach-L methods. This finding confirms that, in the absence of adiposity measures, liver enzyme elevation is the single strongest indicator of fatty liver risk. The subsequent top factors, in descending order of average importance, were: HDL-C (36.82%), UA (24.90%), C-reactive protein (CRP; 21.23%), age (10.05%), FPG (8.17%), and WBC (13.70%). Notably, DBP (4.95%) and LDL-C (6.18%) also demonstrated meaningful importance.

Several observations merit attention. First, the emergence of CRP and WBC as highly important factors in Model 2 highlights the significant role of systemic inflammation in NAFLD pathophysiology when the dominant effect of obesity is removed. Second, HDL-C remained a top-tier factor in both models, underscoring its robust and independent inverse relationship with hepatic fat accumulation. Third, uric acid consistently ranked among the most important predictors, supporting the growing body of evidence linking hyperuricemia to metabolic dysfunction and liver injury. Fourth, it is noteworthy that diastolic blood pressure, but not systolic blood pressure, emerged as a meaningful factor, suggesting that DBP may be more specifically related to the metabolic disturbances underlying NAFLD. Finally, lifestyle factors such as smoking, betel nut chewing, physical activity, and sleep duration showed relatively low importance rankings in both models, indicating that in this healthy young cohort, metabolic and inflammatory markers are more direct determinants of FLI than behavioral variables.

In summary, the machine learning analyses revealed that adiposity (BF) is the dominant predictor of FLI when included in the model. However, when BF is removed, a broader constellation of metabolic (HDL-C, UA, FPG, LDL-C), inflammatory (CRP, WBC), and hepatic (SGPT) factors emerge as key determinants, providing a more nuanced understanding of the multifactorial nature of fatty liver disease in young, healthy men.

Discussion

By using five different Mach-L methods, we identified the most important six factors related to FLI in both Model 1 and Model 2 in healthy Taiwanese young men. Even though there have been several studies done in this field^35,36, our novelties were: (1) Using Mach-L, (2) Using FLI, (3) Focusing on young healthy men without taking medications which might have changed the independent variables. Therefore, we believed that our findings were reliable.

In Model 1, the most important factor was BF. The simple correlation between BF and FLI was 0.745 (p < 0.001). This is not surprising and is in line with many other cornerstone studies^37–40. For example, Fan et al. reported that in overweight and obese subjects, the risk of having NAFLD was 3.55 and 7.59 times higher than in normal weight subjects, highlighting a clear dose-response relationship between BMI and NAFLD risk³⁷. Other studies have further elucidated this relationship, demonstrating strong correlations between body fat percentage, visceral adiposity, and the presence of NAFLD^38,39. The mechanism behind this relationship includes insulin resistance and dietary influences, since insulin resistance can increase lipogenesis in the liver and high-fat diets can enhance hepatic fat accumulation^37,39,40. Our result further supported this relationship.

It is interesting to note that SGPT had a more important impact than SGOT. In the present study, SGPT was the second most important factor. It is well-known that SGPT is an enzyme mainly stored in the liver. When liver cells are damaged or inflamed, SGPT is released into the blood. It is now generally considered that a high SGPT level is a result of liver damage rather than a cause⁴¹. It could also be regarded as a marker for the severity of liver damage. Again, many studies focused on this topic and their results are the same as our result^42–44. As for the importance of SGPT being higher than SGOT, the distribution of SGPT and SGOT could explain this result. SGPT is more liver-specific. Contrary to SGPT, SGOT could be found in many other organs including the liver, heart, kidneys, and muscle^45,46. Thus, it could be concluded that SGPT is a more specific marker for liver disease compared to SGOT.

In our study, HDL-C had a negative relationship with FLI and it was the third most important factor. This highlights that high HDL-C might have a protective role in NAFLD. Xuan et al. pointed out that this relationship could be explained by the role of reverse cholesterol transportation and removal of cholesterol from the liver⁴⁷. Other studies also supported this conclusion^48,49. The relationship between age and NAFLD is important since the prevalence and severity of NAFLD increase with aging. First, from the incidence aspect, Lin et al. divided the cohort into 4 quartiles, from 20 to above 65, each quartile had a 15-year difference in their study. The incidences were 11.7%, 15.9%, 21.5%, and 21.5%, respectively⁵⁰. It is postulated that this relationship might be related to the increased triglyceride and FPG which contribute to the development of NAFLD. At the same time, aging could also change BF distribution, increase insulin resistance, and decline metabolic rate^50,51. The present study not only confirmed the importance of age but also gave priority to aging in the development of NAFLD.

It should be noted that, like MLR, Mach-L could also be thought of as a multiple regression which means that all the impacts of different variables are adjusted. The independent importance of UA was interesting. Our finding was not alone; other studies had the same conclusion. For instance, individuals with elevated UA level have been shown to have an increased chance of NAFLD^52,53. The rationale behind this relationship could be contributed to three factors: (1) Insulin resistance. As aforementioned, insulin resistance could increase lipolysis in adipose tissue and free fatty acids will be accumulated in liver cells^52,54. (2) Inflammation: UA is often regarded as an inflammatory marker which could further exacerbate the damage of liver cells^52,55,56. (3) Higher UA could lead to oxidative stress which also increases liver injury and fat accumulation in the liver^54,56.

From the aforementioned information, the role of IR is repeatedly noted. It is a fact that IR is one of the most important pathophysiologies for developing hyperglycemia⁵⁷. In Model 1, FPG was the 6th most important factor for NAFLD. It is found that, in diabetic patients, higher glycated hemoglobin is related to more severe liver damage. Lopez pointed out that for every 1% point increase in glycated hemoglobin, the likelihood of severe liver fibrosis rose by 15%⁵⁸. The cause behind the role of IR is due to the higher blood glucose which is the substrate for de novo lipogenesis and is deposited in the liver⁵⁹.

Comparing to Model 1, in Model 2, SGPT remained the most important factor, and HDL-C became the second most important factor. The other orders of the factors did not change much. However, CRP and WBC emerged as highly important factors. In short, being inflammatory markers, higher CRP and WBC levels indicate a higher chance of having NAFLD^60,61. Finally, it is also interesting to note that only DBP was more important for NAFLD irrespective of SBP. Qian et al. reported a similar finding to ours that there is an independent relationship between DBP, SBP, and NAFLD. For every unit increase in DBP, the odds of having NAFLD increased significantly⁶². The physiological roles of DBP and SBP are different. SBP is the maximum pressure in the arteries during the contraction of the heart’s ventricles⁶³. At the same time, DBP is the minimum pressure in the arteries when the heart is at rest between beats⁶⁴. However, the details of the independent and complex relationships remain to be solved.

Our study has limitations. First, this is a cross-sectional study which is less persuasive than a longitudinal study. A future study with a longitudinal design will be better to elucidate the importance of these variables to NAFLD. Second, the present study was done in one ethnic group; extrapolation to another ethnic group should be exercised with caution. Third, because this study focused exclusively on mengiven the well-established sex-based differences in body fat distribution, hormonal regulation, and metabolic risk profiles—the findings may not be directly generalizable to women; a parallel analysis of the female cohort is currently underway and will be reported separately.

Conclusion

In the present study, MLR was regarded as a benchmark for comparison. Our findings showed that Mach-L outperformed traditional MLR. By using five Mach-L methods, in Model 1 (with BF), the most important six factors were BF, SGPT, HDL-C, age, UA, and FPG. After removing BF (Model 2), the order changed slightly. They were SGPT, HDL-C, UA, CRP, age, and FPG in a healthy young Taiwanese men cohort.

Abbreviations

NAFLD

Nonalcoholic fatty liver disease

FLI

Fatty liver index

Mach-L

Machine learning

BF

Body fat

SGPT

Serum glutamic pyruvic transaminase

HDL-C

High-density lipoprotein cholesterol

UA

Uric acid

FPG

Fasting plasma glucose

WBC

White blood cell count

DBP

Diastolic blood pressure

MLR

Multiple linear regression

RF

Random Forest

SGB

Stochastic gradient boosting

MARS

Multivariate adaptive regression splines

XGBoost

eXtreme Gradient Boosting

LightGBM

Light Gradient Boosting Machine

SMAPE

Symmetric mean absolute percentage error

RAE

Relative absolute error

RRSE

Root relative squared error

RMSE

Root mean squared error

Author contributions

Interpretation of data, Chiu-Sung Ko; Analysis, Dee Pei; Acquisition, Yu-Chen Tseng; Investigation, Yu-Chen Tseng, Dee Pei, and Ta-Wei Chu; Validation, Yu-Chen Tseng, and Chiu-Sung Ko; Drafted the work– original, Yu-Chen Tseng; Drafted the work – review & editing, Kai-Jo Chiang.

Funding

This study was funded by Taichung Armed Forces General Hospital, grant number TCAFGH_D_113010.

Data availability

The data supporting this study’s findings are obtained from the MJ Health Research Foundation. However, due to access restrictions, these data are not publicly available as they were used under authorization for this study. Nevertheless, they can be requested from the corresponding author upon reasonable request and with approval from the MJ Health Research Foundation.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

The study protocol was approved by Institutional Review Board of the Tri-Service General Hospital (IRB No.:C202405017 and date of approval 22 March 2024). Since there were no sample collection from the patients, a short review of the Institutional Review Board was approved and no consent form was needed.

Informed consent

Not applicable. Analysis was based on secondary data sourced from MJ Health Research Foundation.