New predictive models and indices for screening MAFLD in school-aged overweight/obese children

Currently, most predictions of metabolic-associated fatty liver disease (MAFLD) in school-aged children utilize indicators that usually predict nonalcoholic fatty liver disease (NAFLD). The present study aimed to develop new predictive models and predictors for children with MAFLD, which could enhance the feasibility of MAFLD screening programs in the future. A total of 331 school-aged overweight/obese children were recruited from six primary schools in Ningbo city, China. Hepatic steatosis and fibrosis were detected with controlled attenuation parameter (CAP) and liver stiffness measurement (LSM), respectively. Machine learning methods were adapted to build a set of variables to predict MAFLD in children. Then, the area under the curve (AUC) of multiple models and indices was compared to predict pediatric MAFLD. Compared with non-MAFLD children, children with MAFLD had more obvious metabolic disturbances, as they had higher anthropometric indicators, alanine aminotransferase, fasting plasma glucose, and inflammation indicators (white blood cell count, hemoglobin, neutrophil count) (all P < 0.05). The optimal variables for all subjects selected by random forest (RF) were alanine aminotransferase, uric acid, insulin, and BMI. The logistic regression (LR) model performed best, with AUC values of 0.758 for males and 0.642 for females in predicting MAFLD. LnAI-BMI, LnAI, and LnAL-WHtR were approving indices for predicting pediatric MAFLD in all participants, boys and girls individually. Conclusions: This study developed LR models and sex-specific indices for predicting MAFLD in overweight/obese children that may be useful for widespread screening and identification of children at high risk of MAFLD for early treatment. What is Known: • Most of the indicators predicting pediatric MAFLD are derived from the predictive indicators for NAFLD, but the diagnostic criteria for MAFLD and NAFLD are not exactly the same. • The accuracy of predictors based on routine physical examination and blood biochemical indicators to diagnose MAFLD is limited. What is New: • This study developed indicators based on routine examination parameters that have approving performance for MAFLD, with AUC values exceeding 0.70. What is Known: • Most of the indicators predicting pediatric MAFLD are derived from the predictive indicators for NAFLD, but the diagnostic criteria for MAFLD and NAFLD are not exactly the same. • The accuracy of predictors based on routine physical examination and blood biochemical indicators to diagnose MAFLD is limited. What is New: • This study developed indicators based on routine examination parameters that have approving performance for MAFLD, with AUC values exceeding 0.70.


Introduction
Nonalcoholic fatty liver disease (NAFLD) is a progressive disease ranging from simple hepatic steatosis to nonalcoholic steatohepatitis and cirrhosis.It is closely related to obesity, dyslipidemia, and insulin resistance [1]; affects nearly 10% of the general pediatric population in many countries [2,3]; and is the most common chronic liver disease in children.With comprehensive research on pediatric NAFLD, increasing evidence has indicated that NAFLD is a hepatic manifestation of systemic metabolic disorders [4].
Therefore, an international consensus proposed "metabolic-related fatty liver disease" (MAFLD) as a new name for NAFLD in 2020 [5], and the diagnostic criteria for children were developed in 2021 [6]."MAFLD" is not simply a name change but a more accurate reflection of pathophysiology [7].It helps to link fatty liver disease with other metabolic disturbances and is beneficial for the treatment of fatty liver and supports from the side of patients [8].It is estimated that the prevalence of MAFLD in the general pediatric population is approximately 10% [9], while the prevalence in overweight/obese children is more than 30% [10], suggesting that it is accompanied by a substantial health burden.
Currently, no effective drug has been approved for pediatric MAFLD.Early detection and lifestyle intervention are essential to control the progression of pediatric MAFLD.Liver biopsy is the gold standard for the diagnosis of hepatic steatosis.However, it requires a professional skilled person to perform it under a technical assisting apparatus, which is not easily implemented in the field screening program.Biomarkers, particularly serum lipid levels and aminotransferase, are widely used in the screening of hepatic steatosis.However, the accuracy and predictive value were generally low.To date, most current predictors for MAFLD have utilized the predictors for pediatric NAFLD, for example, the triglyceride and glucose index (TyG), TyG-BMI [11], and pediatric NAFLD fibrosis index (PNFI) [12].Notably, the diagnostic criteria for pediatric MAFLD are not the same as those for NAFLD and some lean NAFLD children cannot be categorized as MAFLD children and vice versa [13], suggesting the necessity in developing new less invasive indicators for screening MAFLD in overweight/obese children.
Therefore, in the present study, several machine learning methods were adopted to seek for the proper biomarkers (anthropometric parameters and readily used clinical laboratory indicators) and generate predictive models to screen for MAFLD in overweigh/obese children.

Study design and population selection
The data of the present study were extracted from children participating in the Healthy Weight Management Project for Children and Adolescents in Ningbo city, Zhejiang province, China (RCT Registration No.: NCT05482191).All overweight/obese children in third grade across six primary schools in the Haishu, Yinzhou, and Zhenhai districts of Ningbo city were recruited in the present study.The project began in September 2022, and the intervention was expected to last for one academic year, which aimed to explore and establish effective interventions for children.The baseline information of all overweight/obese children was used in present study.The project was performed in line with the principles of the Declaration of Helsinki and was approved by the Ethics Committee of Ningbo First Hospital (Approval No. 2021-R168).Informed consent was obtained from all participants and their parents.
Participants without enough information to be categorized as MAFLD patients were excluded.Specific exclusion criteria were lacking qualified results of vibration-controlled transient elastography (VCTE) (n = 19), absence of important laboratory covariates (n = 32), and children with normal weight (n = 40).Finally, 331 participants were included in the present study (213 boys, 118 girls; Fig. 1).

Definition of hepatic steatosis and fibrosis
Controlled attenuation parameter (CAP) and liver stiffness measurement (LSM) were performed with FibroScan Handy (Echosens, Paris, France), which was equipped with a medium probe.One sonographic technician from Ningbo First Hospital who received standardized training conducted the examinations on the liver.Participants with eligible FibroScan tests (≥ 10 successful stiffness measures and liver stiffness interquartile range/ median < 30%) were included in the current analysis.Hepatic steatosis and fibrosis were diagnosed by CAP ≥ 248 dBm [14] and LSM ≥ 5.1 kPa [15], respectively.

Definition of MAFLD
The diagnosis of MAFLD was based on hepatic steatosis or fibrosis, and at least one of the following conditions: (1) excess adiposity (overweight/obesity/abdominal obesity), (2) prediabetes or type 2 diabetes, and (3) metabolic dysregulation [6].Overweight/obesity was defined as BMI z scores > 1 SD for children aged 5-10 years according to the WHO growth reference [16], and abdominal obesity was defined as waist circumference (WC) > 90th percentile.Since no one is fasting plasma glucose exceeded 7.0 mmol/L, prediabetes was defined as fasting plasma glucose ≥ 5.6 mmol/L [17].For children between 8 and 9 years old, metabolic dysregulation was defined as the presence of at least two of the following abnormal conditions: (1) systolic blood pressure (SBP) or diastolic blood pressure (DBP) > 90th percentile, (2) plasma triglyceride concentration (TG) > 90th percentile, (3) plasma high-density lipoprotein cholesterol (HDL-C) ≤ 10th percentile, and (4) triglyceride-to-HDL cholesterol ratio > 2.25.The cut-off values for WC, lipid profile, and blood pressure were selected according to Chinese age-sex-specific criteria [18][19][20].

Anthropometric and blood measurements
Anthropometric measures were performed by trained health workers from the community healthcare centers according to the standard protocols.Height was measured to 0.1 cm with a mechanical height meter, and weight was measured by a bodycomposition instrument (Inbody770, Biospace, California, USA) to 0.01 kg with participants removed their coat and shoes.BMI was calculated as weight divided by square of height (kg/ m 2 ).WC was measured as the circumference of the midpoint line between the lowest costal point and the upper edge of the iliac crest, and hip circumference (HC) was measured as the horizontal circumference of the most prominent part behind the buttocks.Both were nearest to 0.1 cm.The measurements of SBP and DBP followed the manufacturer's instructions.Blood samples were collected from participants who fasted for at least an 8 h.The waist-to-height ratio (WHtR) was calculated as WC (cm)/height (cm), and WHR was calculated as WC (cm)/ HC (cm).Routine blood tests including lymphocyte absolute count (LYMPH), basophil absolute count (BASO), white blood cell count (WBC), eosinophil count (EO), hemoglobin (HGB), platelet count (PLT), and neutrophil count (NEUT) were measured by automated hematology analyzers.Alanine aminotransferase (ALT) was measured by an instrument (Beckman Coulter AU680, Brea, USA) using an enzymatic method.Insulin was measured by an autoanalyzer (Roche Cobas e602, Mannheim, Germany).FPG, low-density lipoprotein cholesterol (LDL-C), HDL-C, total cholesterol (TC), TG, creatinine (Cr), estimated glomerular filtration rate (eGFR), blood uric nitrogen (BUN), and uric acid (UA) were measured by another autoanalyzer (Beckman Coulter AU5800, Brea, USA).

Indicators used for comparisons
Two effective predictors of NAFLD, the TyG index and homeostasis model assessment of insulin resistance (HOMA-IR), were selected as positive references because they had good performance in the prediction of pediatric MAFLD [11].
The grade I indices were constructed with a similar equation as TyG, which is listed below: Index Y = candidate variable A ^ candidate variable B = Ln [variable A × variable B/2].As previous study indicated that TyG-BMI had the highest area under the curve (AUC) in predicting NAFLD in children [11], the selected promising grade I indices were further combined with obesity-related anthropometric variables to construct grade II indices, for instance:

Statistical analyses
All analyses were performed by R version 4.3.1.All 26 variables (shown in Table 1) were included in the random forest (RF) model to obtain the importance ranking of the variables.The variables were sequentially included in the random forest model combined with the leave-one-out method.According to the ranking order and the AUC value, high priority variables were selected to construct models for screening MAFLD.Four machine learning models including RF (R package: "randomForest"), support vector machine (SVM, R package: "pROC"), logistic regression (LR, R package: "e1071"), and decision tree (DT, R package: "rpart") were used to validate the separating capacity of the selected variables in different models, and their total performance was judged by the AUC as well.
Based on the best combination of variables selected by RF, the variable with the highest importance was combined with other variables to construct grade I indices according to the calculation form of the TyG index.Then, the grade II indices were further combined with obesity-related anthropometric indices (WC, BMI, WHR, and WHtR) following the format of grade II index construction.
The sensitivity and specificity of the indices were calculated according to the cut-off value at the maximum Youden index.The bootstrap method was used for pairwise comparisons between the AUCs of the indices.A two-tailed p value < 0.05 was considered statistically significant.

Results
In the total population, the mean age was 8.56 ± 0.30 years, and 64.35% were boys.A total of 4.83% of the children had prediabetes, and 23.87% had hypertension.There were 65 children diagnosed with MAFLD, accounting for 19.64%.Compared to children without MAFLD, children with MAFLD had higher anthropometric measurements including BMI and WC, higher CAP and LSM values, higher levels of serum insulin, FPG, TG, ALT and UA, and higher chronic inflammation indicators (WBC, HGB, NEUT) (all P < 0.05; Table 1).In addition, the results of the total population remained similar in boys (Table S1).In terms of girls, MAFLD participants have higher SBP, higher LDL-C levels, and a higher prevalence of hypertension than those without MAFLD (P < 0.05; Table S1).

High-ranking variable selection for predicting MAFLD in children
The importance ranking of the variables obtained by the RF in all participants is shown in Fig. 2A, where higher importance means a greater contribution to distinguishing MAFLD from non-MAFLD in children.Figure 2B shows the changes in the AUC of the RF model to predict MAFLD in all subjects with increasing included variables.The AUC of the top four variables (ALT, UA, insulin, and BMI) was 0.704, and when the fifth variable was added, the AUC suddenly decreased.Therefore, we selected the top four variables for the construction of the grade I indices.Using the same method, an important ranking of variables was performed in both boys and girls (Fig. 2C and E), individually, and the top five and seven variables were included for subsequent analyses (Fig. 2D and F).

Model construction for predicting MAFLD in children
The analysis results of the four models, including LR, RF, SVM, and DT, for predicting MAFLD in all subjects, boys and girls, are shown in Fig. 3 and Table S2.RF was the best model for predicting MAFLD in children, with an AUC of 0.703.For sex-specific predictions, the LR model performed best in predicting MAFLD, with AUC values of 0.758 and 0.642 for males and females, and NPV values of 0.919 and 0.927, respectively.The specific forms of the LR models for boys and girls are " − 14.976 + 0.054 × ALT + 0.013 × Insulin + 0.169 × BMI + 0.003 × UA + 1.363 × FPG" and " − 5.173 + 0.058 × ALT-1.140× Age + 0.002 × Insulin + 0.514 × LDL + 0.05 7 × S BP + 0. 008 × BMI + 0.068 × W C," re spe ctively.

Grade I index construction and evaluation
The positive predictive value (PPV), negative predictive value (NPV), AUCs, and 95% CI of the grade I indices are shown in Table 2.All indices could significantly predict MAFLD in all subjects and boys (all P < 0.05).In all participants and in boys, the ALT^Insulin (abbreviated as LnAI) index showed the highest AUCs (0.703 and 0.739) and NPVs (0.910 and 0.929), which meant that this index can identify children without MAFLD well.ALT^TG had significantly higher AUC values than TyG (P < 0.05; Tables S3 and S4).In girls, ALT^LDL showed the largest AUC of 0.680 for predicting MAFLD, but it was not significantly higher than TyG or HOMA-IR indices (P > 0.05; Table S5), with PPV and NPVs of 0.309 and 0.905, respectively.

Grade II index construction and evaluation
The comparisons of the performance of grade II indices for predicting MAFLD are shown in Table 3.All grade II indices significantly predicted MAFLD (all P < 0.05).In all participants, LnAI-BMI had an AUC value of 0.712 and the highest NPV of 0.881.Its AUC value, PPV, NPV, and diagnostic odds ratio (DOR) were higher than those of the TyG index and all its modified indices, but the differences were not significant (all P > 0.05).The AUCs of LnAT and its modified indices (AUC: 0.694-0.712)were significantly higher than those of the TyG index (all P < 0.05; Table S6), and their NPVs were higher than those of TyG and its modified indices.In boys, the AUC values of the modified LnAI indices were lower than those of LnAI.LnAI and LnAT-BMI had the same largest AUC of 0.739 for predicting MAFLD, among which the former had the largest NPV and DOR of 0.929 and 6.60, respectively, and the latter had a significantly better AUC than the TyG index (P < 0.05; Table S7).In girls, The larger the ordinate is, the higher the importance of the variable.B, D, and F The AUC values of the random forest models for predicting MAFLD when different variables were cumulatively included in the model for the total population, boys and girls, respectively LnAL and its modified indices had good performance for predicting MAFLD as their AUC values were higher than those of TyG and its modified indices, but the differences were not significant (all P > 0.05; Table S8).LnAL-WHtR had the best predictive performance as it had the highest AUC value (0.716).

Discussion
The definition of pediatric MAFLD is not equivalent to that of NAFLD.The need to generate proper predictors for screening MAFLD is urgent.Thus, our study explored new predictors for screening MAFLD in children with machine learning methods combined with formula construction.Among the four machine learning methods, LR performed best in boys and girls with AUC values of 0.758 and 0.642, respectively.Among those newly generated predictors, LnAI-BMI, LnAI, and LnAL-WHtR showed approving performance for predicting MAFLD in all participants, boys and girls, respectively.
Liver biopsy is the gold standard for the diagnosis of fatty liver, but this technique is invasive and expensive, with sampling and diagnostic errors.Therefore, it is important to evaluate hepatic steatosis using less invasive methods and readily available parameters.Among the alternative methods to liver biopsy, VCTE can quantify the level of hepatic steatosis and fibrosis and has shown good accuracy in the diagnosis of NAFLD [22].However, the instrument requires professional technicians to operate and is not involved in the school physical examination or routine examination of children.In contrast, the predictive parameters constructed by blood examination indicators and physical examination indicators in this study are easier to obtain, which greatly reduces the labor and time costs and is suitable for largescale promotion.
LR performed better than the other three models in predicting MAFLD in boys and girls in this study.Machine learning algorithms can build complex models to make accurate disease predictions [23], but the internal structure and decision-making process of complex models such as RF and SVM are difficult to explain.In contrast, LR is highly interpretable and has good clinical practicability.
When screening variables in total participants and different sexes to predict MAFLD, ALT showed critical function, which might be attributed to its physiological role.ALT, an intracellular functional enzyme that catalyzes the transfer reaction of amino acids [24], is present at the highest concentration in hepatocytes and at very low concentrations in any other tissue.When hepatocytes are damaged, it can cause an increase in serum ALT levels.Therefore, serum ALT levels can reflect the process of liver injury, including fatty liver disease and viral-related liver disease [25].In this study, ALT significantly predicted MAFLD in children, with AUC values ranging from 0.603 to 0.704.Although the ALT level can be used as an important reference, it cannot be directly applied as a diagnostic criterion for fatty liver, as some patients with normal ALT levels may have liverrelated disease [26].A meta-analysis of 11 studies showed that approximately 25% of individuals with NAFLD and 19% of individuals with NASH had normal ALT levels [27].Therefore, it is necessary to combine ALT with other related parameters to form a composite index to improve the accuracy of screening.
The accumulation of TG in the cytoplasm of hepatocytes is the main feature of hepatic steatosis [28].Under physiological conditions, the balance between the synthesis and consumption of TG results in low steady-state TG concentrations within hepatocytes [29].However, excessive fatty acid intake or abnormal fatty acid synthesis can lead to the accumulation of TG in the liver [28].In addition, insulin plays a role in inhibiting glucose production and promoting fatty acid synthesis in the liver.Hepatic insulin resistance reduces the suppression of glucose production, but it still promotes de novo lipogenesis and TG accumulation in the liver [30].Insulin resistance and TG accumulation in the liver contribute to the formation of very low-density lipoprotein (VLDL) and small dense LDL [31].Most of the VLDL will be catabolized to LDL, resulting in an increase in serum LDL concentration.ALT alanine aminotransferase, AUC area under the curve, BMI body mass index, SBP systolic blood pressure, FPG fasting plasma glucose, HDL high-density lipoprotein cholesterol, HOMA-IR homeostasis model assessment of insulin resistance, LDL low-density lipoprotein cholesterol, PPV positive predictive value, NPV negative predictive value (NPV), TC total cholesterol, TG triglyceride, TyG triacylglycerol-glucose, UA uric acid, WC waist circumference, WHR waist-to-hip ratio, WHtR waist to height ratio a P determined using the receiver operating characteristic curve for AUC  Obesity is another important risk factor for MAFLD in children and increases the incidence of many diseases, including diabetes, hypertension, and dyslipidemia.Matsuda and Shimomura [32] declared that obesity may be linked to these related diseases through oxidative stress.When combined with obesity-related anthropometric indices, most of the modified indices developed in this study showed better performance than the original indices.BMI measures body fat using weight and height, but the weight used for the calculation cannot distinguish between fat and lean mass [33].In contrast, WC, WHR, and WHtR are better indicators of visceral adiposity than BMI and have been considered proxies for abdominal obesity in a previous study [34].Therefore, WC, WHtR, and WHR were suspected to predict MAFLD better than BMI when combined with the new indices and TyG index.However, in all participants and boys in our study and in other similar studies [11,35], some indices combined with BMI performed better than other obesity-related indices.Abdominal fat included visceral fat and subcutaneous fat, with visceral fat being a better predictor of fatty liver than subcutaneous fat [36].However, WC, WHtR, and WHR could not distinguish between visceral fat and subcutaneous fat [37].In addition, WC, which excludes height, estimated lower abdominal fat mass for short people than for tall people.
An interesting result in this study was that in all participants and in boys, although the AUC value of LnAI was higher than that of LnAT, only the AUC of LnAT was significantly higher than that of the TyG index.One reason for this was the small number of participants in our study population and the results were susceptible to abnormal insulin values.The other was that the study population in our study was overweight/obese children aged 8-9 years with high and wide ranges of insulin levels.They may contribute to the nonsignificant results.In addition, the AUC values of modified LnAI indices were lower than LnAI in boys, which may be related to the wide range of insulin in boys.The mean value of insulin in boys with MAFLD was approximately 1.6 times that of non-MAFLD, but there was no significant difference in girls.This also explained why some of the modified LnAI indices had higher AUC values than LnAI in all participants.
The present study is the first to explore appropriate and simple indicators with an adequate amount of individual data to predict MAFLD in children.Models and new indices were developed based on biochemical data and anthropometric data, some of which could be collected from the Chinese National Survey on Students' Constitution and Health, supported by national finance.From the cost-effect aspect, a simplified LnAI or LnAI-WHR indicator would only add the ALT, insulin, and TG test in the overweight/obese children, which enhanced the feasibility and could also warn that children should pay attention to the impact of obesity.However, there are some limitations.First, our study included only overweight/obese children, which could not be extrapolated to the whole population or the lean MAFLD population.Second, our study had a limited sample size, but it was well represented, as the study population came from six primary schools in different areas.Finally, CAP and VCTE were applied in the diagnosis of MAFLD in this study, which still falls short compared with liver biopsy.However, CAP and VCTE are widely used for MAFLD diagnosis.

Conclusion
The present study established LR-based prediction models for MAFLD in overweight/obese children and explored new predictors.LnAI-BMI, LnAI, and LnAL-WHtR were satisfactory indices for predicting pediatric MAFLD in all participants, boys and girls.All had better predictive performance than the TyG index.In addition, routine physical measurements and laboratory tests can only obtain ALT levels and obesity-related indicators, so blood lipid testing should be added to overweight/obese children to construct indicators for widespread screening and identification of children in need of liver ultrasound and early treatment on a large scale.

Fig. 1
Fig. 1 Flow chart for the selection of study subjects.After screening, 331 subjects were enrolled and divided into MAFLD and non-MAFLD groups

Fig. 2
Fig. 2 Importance ranking of variables and the cumulative AUC of the candidate biomarkers for predicting MAFLD in total participants, boys and girls.A, C, and E The ranking of importance of variables in predicting MAFLD in all populations, boys and girls, respectively.

Fig. 3
Fig. 3 The ROC curves of the candidate biomarkers for the diagnosis of MAFLD using RF, LR, SVM, and DT methods.This figure presents the ROC curves of the four models (RF, LR, SVM, and DT)

Table 1
Baseline characteristics of subjectsVariables are shown as n (%) or mean (SD) ALT alanine aminotransferase, BASO basophil absolute count, BMI body mass index, BUN blood uric nitrogen, Cr creatinine, CAP controlled attenuation parameter, DBP diastolic blood pressure, eGFR estimated glomerular filtration rate, EO eosinophil count, FPG fasting plasma glucose, HC hip circumference, HDL-C high-density lipoprotein cholesterol, HGB hemoglobin, LDL-C low-density lipoprotein cholesterol, LSM liver stiffness measurement, LYMPH lymphocyte absolute count, NEUT neutrophil count, PLT plate- let count, SBP systolic blood pressure, TG triglyceride, TC total cholesterol, UA uric acid, WC waist circumference, WBC white blood cell count, WHR waist to hip ratio, WHtR waist to height ratio a Pearson's chi-squared test for categorical variables b Student's t test c The Kruskal-Wallis test for continuous variables

Table 2
Cut-off values and AUCs of grade I indices for predicting MAFLD

Table 3
Cut-off values and AUCs of grade II indices for predicting MAFLDAUC area under the curve, TyG triacylglycerol-glucose, BMI body mass index, DOR diagnostic odds ratio, PPV positive predictive value, NPV negative predictive value, WC waist circumference, WHR waist-to-hip ratio, WHtR waist-to-height ratio a P determined using the receiver operating characteristic curve for AUC