JMIR Publications

ABSTRACT

Background:

The morbidity of nonalcoholic fatty liver disease (NAFLD) is increasing in China, and simultaneously it has become a huge health burden. Traditional ultrasound, as a screening tool for fatty liver, is unable to quantitative steatosis, resulting in a large number of patients with moderate-severe steatosis cannot be followed up. Transient elastography (TE) allows quantitative diagnosis of steatosis and fibrosis, with higher concordance with biopsy. Machine learning (ML) technology is used to construct diagnostic models for NAFLD using a variety of laboratory indicators.

Objective:

We hope to derive a new diagnostic model that can be used for staging hepatic steatosis through ML using TE results as a criterion. And we simplify the input features of the model to obtain an inexpensive and easy-to-use model to differentiate patients with NAFLD who need follow-up.

Methods:

We screened the data of health examination at Suzhou Municipal Hospital during March-May 2023, finding 978 residents who had TE and complete medical records. Classification models such as logistic regression (LR), k-nearest neighbor (KNN), support vertical machine (SVM), random forest (RF), light gradient boosting machine (LightGBM), and extreme gradient boosting (XGBoost) were developed to predict NAFLD with different severity. The performance of the six models was evaluated by area under the receiver operating characteristic (AUROC).

Results:

A total of 916 patients were included in this study; of those 273 patients had moderate-severe NAFLD. The concordance of traditional ultrasound and TE for the detection of moderate-severe NAFLD, were 84.6%. The AUROC of RF, LightGBM, XGBoost, SVM, KNN and LR was 0.91,0.86,0.83,0.88,0.77and0.81 respectively. Additionally, the accuracy of above models was 84%,81%,78%,81%,76% and 77%.RF had the best performance. We simplified a cheap and easy-to-use RF model (AUROC 0.88) which contains 6 features including Waist Circumference,BMI,FPG,UA,TBil and hs-CRP.

Conclusions:

We propose a cheap algorithm through ML that can be used to identify moderate-severe NAFLD by screening data from health examination in order to clarify their need for special investigation and intervention.