Non-alcoholic fatty liver disease (NAFLD) has become a sever public health problem worldwide (1, 2). The prevalence rate of NAFLD is around 20~30% and is increasing constantly. In the past 10 years, the prevalence rate of NAFLD has doubled (3). NAFLD is closely related to overweight or obesity, hyperlipidemia, type 2 diabetes mellitus (T2DM) and other chronic metabolic diseases: the prevalence of NAFLD is 60–90%, 27–92%, and 28–70% in obesity, hyperlipidemia and T2DM, respectively (4). NAFLD is a group of disease spectrum, the development of which is liver steatohepatitis, non-alcoholic steatohepatitis (NASH), liver fibrosis, cirrhosis, and even liver cancer. NAFLD is main cause leading to the fastest growing of liver cancer, and NASH has become the leading cause of liver failure in the United States (5–8). In recent years, the prevalence of NAFLD in China has gradually increased, and the prevalence has become younger: in 2014, a large sample meta-analysis reported that the prevalence of NAFLD in adults in mainland China was 20.09% (9). Therefore, the large-scale cohort or epidemiological study of NAFLD is of great significance. The implementation of the national physical examination encourages large-scale research, but a simple and easy method is still needed to classify NAFLD patients in the population.

Histologic biopsy is the gold standard for diagnosis of NAFLD, but it is invasive and requires high technology. Ultrasound, CT and MRI are the common diagnostic methods, but the cost of imaging examination is high when large-scale population screening. In order to facilitate the diagnosis of NAFLD, several predictive models have been introduced. Fatty liver index is an algorithm based on serum triglyceride and gamma glutamyl transferase (GGT) levels, body mass index (BMI) and waist circumference, which can predict liver steatosis in general population (10, 11). NAFLD liver fat score uses a formula including metabolic syndrome, T2DM, fasting serum insulin, aspartate aminotransferase (AST) and alanine aminotransferase (ALT) levels to estimate the percentage of liver fat content (12). SteatoTest is a logistic regression model of 12 predicting parameters: a2-macroglo-bulin (A2M), apolipoprotein A1 (ApoA1), haptoglobin, total bilirubin, GGT levels, cholesterol, triglycerides, glucose, age, gender and BMI (13). A prediction model based on laboratory includes six parameters: alanine aminotransferase, high-density lipoprotein cholesterol, triglyceride, hemoglobin A1c (HbA1c), white blood cell count and the presence of hypertension, and this model is used for the screening of NAFLD in common population (14). However, there is a problem that these prediction parameters are difficult to obtain. Although the existing NAFLD prediction models have been widely used, their application in large-scale epidemiological research and many areas of developing countries like China is limited.

It has been applied in medicine to establish accurate prediction model through machine learning. Machine learning outperforms conventional statistical methods with its ability to better identify variables relevant to clinical outcomes, better predictive performance, better modeling of complex relationships, ability to learn from multiple modules of data, and robustness to data noise. These tools have been used to diagnose fatty liver, meningitis, glaucoma, coronary heart disease, cancer and other diseases (14–21). Our purpose is to use machine learning to analyze the data of 304,145 physical examinees, and to establish a simple NAFLD screening model that does not rely on indicators tested in laboratory.

Non-alcoholic fatty liver disease (NAFLD) is the most common liver disease in the world, which is the main cause of liver cirrhosis and liver cancer. NAFLD lacks effective drug treatment, so early identification of disease and early prevention have the most effective means to improve the disease. In this study, through 12 questionnaires and physical measurement variables, four ML screening models based on 304,145 subjects for NAFLD in large-scale physical examination population were established, XGBoost got the best performance in the validation, which had accuracy = 0.880, precision = 0.801, recall = 0.894, F-1 = 0.882 and AUC = 0.951.

Detailed analysis of existing epidemiological data shows that the risk factors of NAFLD in China are similar to those in the West and other parts of Asia, metabolic syndrome (MetS) is associated with higher risk of non-alcoholic steatohepatitis and more progressive disease 0.2, In this study, BMI, waist, hypertension, gallbladder disease and T2DM were all the risk factors of MetS (40–42). On one hand, MetS is a strong predictor of NAFLD, while on the other hand, NAFLD is a good predictor for the clustering of components of MetS (43). In addition, a number of other risk factors for NAFLD have been identified in Chinese studies. These risk factors include advancing age, male gender, physical inactivity, high-fat intake, high-sugar intake, overeating, smoking, expanding waist circumference, and high-raising BMI (41, 42, 44). The conclusions of these studies were consistent with those of this study.

Our research has several advantages. First of all, some of the existing NAFLD prediction models involve laboratory and clinical parameters, and obtaining these parameters requires high human and financial resources, which limits the application of these models in large-scale epidemiological research and areas with poor health care level (10, 14, 45). All the variables in this study come from non-invasive and easily available measurement indicators and questionnaire indicators. This model can be applied to the prediction of NAFLD in the early stage and non-invasive, without expensive laboratory tests, especially in the areas with high epidemiological risk and low socio-economic status.

Secondly, this study is based on a large number of Chinese populations, which has a wide range of choices and is more extrapolated and representative. In addition, our data set covers many major ethnic groups in China, thus better assessing the characteristics of China's population.

Third, the occurrence and development of NAFLD are closely related to lifestyle, so improving lifestyle is an effective treatment (46). Our model not only can be used as a screening model for NAFLD, but also includes adjustable indicators such as diet, smoking, exercise, etc., which can guide people to prevent and delay the occurrence of disease through a healthy lifestyle. Although it is not clear whether exercise has independent benefits for NAFLD, exercise do can improve cardiovascular health, reduce weight, reduce peripheral, fat and liver insulin resistance.

Fourth, the analysis of NAFLD data is a challenging issue because most of the medical data are nonlinear, non-normal, correlation structured, and complex in nature. This study used LASSO penalized logistic regression vs. ML algorithms. LASSO works by shrinking the estimates of the regression coefficients and prevent overfitting due to collinearity of the covariates, which combines the advantages of selection process (easy to explain) and expression (robust), which is particularly useful in large data sets requiring efficient and fast algorithms (47–49). ML algorithms' outstanding performance in the field of processing complex data structures and big data makes it dominant in the field of healthcare and medical imaging, and compared with other machine learning methods, the performance of XGBoost can be improved more than 10 times (25, 50–53).

Surprisingly, compared with patients having non-NAFLD in previous studies, patients tend to eat a high calorie diet, especially in the form of carbohydrates and fats. Zelber Sagi et al. showed that NAFLD patients consumed more soft drinks and meat than the control group (54). Soft drinks contain a lot of sugar, and the intake of sugar is related to NAFLD (55). Musso et al. found that NAFLD patients had higher levels of saturated fat and cholesterol and lower levels of unsaturated fatty acids in their diet than healthy people (56). Although the ideal diet for NAFLD patients has not been determined, the data indicate that diet is important (57). However, in our study, we only got the weak effect of meat and vegetable combination, salt and oil preference on NAFLD (Figure 3), but not the effect of sugar preference on NAFLD. A possible reason for the irrelevance may be that the NPE diet survey was a cross-sectional study, with no professional evaluating the diet of the examined population. The main reason for the errors was that the self-reported eating habits of people undergoing physical examination were highly subjective and lack of professional evaluation indicators. Therefore, more accurate results can be obtained through follow-up of people's lives in future studies. Several limitations existed: firstly, previous studies confirmed that education and family history were important determinants of NAFLD, but we failed to obtain the education and family history of participants. Secondly, we lacked of objective and unified evaluation standard for some indicators, such as dietary status, which may reduce the accuracy of the prediction model. Thirdly, the data used in this study was the physical examination data of China, which might limit the extrapolation of the results. However, this study, based on a large sample of government, is one of the few literature studies providing NAFLD comprehensive epidemiological data for model development. Finally, the parameters in the dataset are not enough to compare with the scores of existing NAFLD prediction models. However, the purpose of this study is to provide a convenient and easily accessible model for the diagnosis of NAFLD through questionnaires and physical measurement. The results show that our model has high diagnostic accuracy and prediction ability.

This study used a simple NAFLD screening model based on a large sample of 304,145 Chinese. The model can obtain high accuracy without relying on laboratory measurement parameters, especially in areas with poor economic conditions and high epidemiology.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

This study was performed in accordance with the principles outlined in the Declaration of Helsinki and approved by the Xinjiang Uygur Autonomous Region CDC Ethical Committee and the Institutional Review Board. People who signed a written informed consent were eligible to participate in the study.

MX and HY conceived the study. YZ and YW collected the data. MX and WJ performed the statistical analyses and drafted the manuscript. HY critically reviewed and edited the manuscript. All authors contributed to data analysis, drafting and revising the article, gave final approval of the version to be published, have agreed on the journal to which the article has been submitted, and agree to be accountable for all aspects of the work.

This work was supported by the Region Social Science Foundation of Xinjiang (Grant No. 2021D01C238).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.