A machine learning classifier for screening non-alcoholic fatty liver disease in the general population

In this study, decision trees, RF, XGBoost, and SVM were used to develop classification models to screen subjects for NAFLD from asymptomatic and general adults. All models were evaluated by accuracy, PPV, F1 score, AUROC and AUPRC. The best performance was shown by the model based on SVM, followed by RF.

Our results showed that the model with SVM was the best model in performance, followed by the RF model. The RF model has shown good performance in previous studies on NAFLD screening^FiveSVM has shown excellent performance in developing medical models involving the classification of NAFLD patients.^Five,6,8Our results were therefore consistent with previous studies^Five,6.

Moreover, our results showed that RF and XGBoost models outperform decision trees, especially RF models. Since the RF and XGBoost models are an integrated decision tree,^9,Ten, these models performed better than single decision tree models. In our research, the RF model performed slightly better than his XGBoost model. This may be partially related to the data not being perfectly balanced between the two groups (non-NAFLD group vs NAFLD group = 10,028 vs 4411). RF algorithms can balance errors in imbalanced data and get good results⁹Still, the performance of models from XGBoost was very close to that of RF, and most of the performance evaluation results were within 0.01 between them (Table 2). Furthermore, because our model is based on the prevalence of NAFLD in the real world, even without a perfectly balanced sample, the model is more robust than models from intentionally designed samples in the real world. may be applicable.

Model performance in accuracy, PPV, AUROC and AUPRC were similar to those reported in previous studies^Five,6 (Supplementary table 1). Comparing with the model using the same algorithm in Liu’s study, his PPV for the model using XGBoost and SVM was similar to that of the previous study.⁶ (PPV: XGBoost: 0.778 vs 0.806, SVM: 0.792 vs 0.768). For AUROC, his XGBoost model in the previous model performed slightly better than ours, but his SVM model in both studies is similar (XGBoost: 0.833 vs 0.873; SVM: 0.850 vs 0.865). For AUPRC, the model with his XGBoost and his SVM in Liu’s report performed slightly better than ours.⁶ (XGBoost: 0.704 vs 0.810; SVM: 0.712 vs 0.800). The accuracy of the model using RF and SVM was slightly lower than the previous Ma report^Five (RF: 0.789 vs 0.827, SVM: 0.798 vs 0.827). In contrast, our model showed slightly higher F1 scores than Ma’s and Liu’s reports.^Five,6 (Ma’s report and comparison: Decision Tree: 0.764 vs 0.569; RF: 0.782 vs 0.579; SVM: 0.792 vs 0.557. Liu’s report and comparison: XGBoost: 0.779 vs 0.695; SVM: 0.792 vs 0.713). The F1 score is a better indicator than the accuracy score for evaluating model accuracy using data that are not absolutely balanced in both groups. Therefore, our model performed similarly to the previous model.

Previous research⁶ Physical examination, complete blood count, liver function tests, lipid panel, renal function tests, and features of tumor markers were included. Whether or not it is closer to NAFLD, more features may improve the model’s performance. However, more features and more complex models can be more difficult to scale and apply to primary care physicians.In contrast, model functionality was simpler and included physical examination, complete blood count, liver function test and lipid panel results (Supplementary Table 2). Therefore, our model has more application in primary care practice and may have similar performance when compared to previous ones.

From our model, the most important features were BMI, TG, VLDL-C, ALT levels, and blood AST/ALT ratio.Our results were consistent with those of previous studies^Five,6BMI and TG are risk factors for NAFLD in adults^11,12, identified for all models. This may be related to the aberrant regulation of hepatic TG accumulation via de novo lipogenesis in NAFLD.¹³Additionally, the liver secretes TG in the form of VLDL for delivery to peripheral tissues. VLDL-C overproduction is one of the hallmarks of NAFLD.¹³This reflects increased de novo lipogenesis and lipolysis of intrahepatic and intraperitoneal fat in NAFLD.¹⁴Furthermore, in nonobese Chinese, increased ALT/AST ratio is associated with risk of new-onset NAFLD.¹⁵The decreased ratio of AST/ALT in NAFLD in our study was consistent with previous findings¹⁵There are many possible reasons for this. For example, a decreased AST/ALT ratio is associated with chronic liver inflammation, insulin resistance, or hepatic steatosis, leading to NAFLD.^16,17Therefore, our study’s model identified the importance of the AST/ALT ratio in BMI, TG, VLDL-C, ALT, and NAFLD. Furthermore, we suggested that these features are important in screening common and asymptomatic adults for her NAFLD.

In our study, we acknowledged that the diagnosis of NAFLD was based on abdominal ultrasound rather than liver biopsy. Abdominal ultrasound is widely accepted as an accurate, non-invasive tool in the diagnosis of NAFLD.¹⁸In our study, the prevalence of NAFLD was 30.5% (4411/14,439). This is similar to China’s national prevalence of 29.2% based on the database from 2008 to 2018, which has increased over the years.¹⁹Diagnosis of NAFLD by abdominal ultrasound may therefore be accurate and reliable in our study. Additionally, a history of type 2 diabetes (T2D) was not included in the candidate features.This is because we cannot diagnose her T2D from a subject based on data, and her privately reported medical history cannot rule out bias, since T2D is a risk factor for her NAFLD.^11,12the model should perform better if T2D is included as a candidate feature.

In conclusion, a classification model can be developed using machine learning based on annual health examination data to screen general adults for NAFLD without duplicate testing. The model with SVM is the best, followed by the model with RF. These models may provide a readily available tool for physicians and primary care physicians to screen for her NAFLD in the general population, and may benefit her NAFLD patients from early diagnosis. .