In this multicenter study, we designed a prediction model based on ML to accurately assess the liver fibrosis stage in CHB patients. Compared with traditional statistical models such as APRI and FIB-4, the ML model showed significant improvement and was easy to process. This suggests that ML has great potential in the field of non-invasive liver fibrosis assessment. Furthermore, our study results showed that the ML model provided similar diagnostic efficacy as the reference standard liver biopsy, which may provide a reliable rationale for the further development of simple, easy-to-use and accurate tools to assess liver fibrosis.
In this study, we used ML methods to more accurately assess the stage of liver fibrosis and improve the accuracy of our model. The final results revealed that our model showed superior accuracy compared with traditional serological models such as APRI and FIB-4. It was also significantly higher than the diagnostic efficacy of 17 non-invasive liver fibrosis models in Chinese patients with hepatitis B mentioned in the study by Li et al.19Furthermore, stratified analysis in inflammation subgroups was performed, and the results showed that it did not significantly affect the performance of the ML model. These findings suggest that ML models may overcome the effects of inflammation in liver cirrhosis assessment, which may be a potential breakthrough in non-invasive diagnosis. This was helped by a novel approach to model building with the following main advantages: first, we compared the performance of models built with several ML methods, and then we focused on and validated the DT model, which has better performance and is easy to use. Indeed, the DT model has been applied to the assessment of liver fibrosis in hepatitis C and has shown significant performance.20Furthermore, previous studies have mainly used classification methods (logistic regression analysis).twenty oneIn most patients, features were selected by univariate tests (e.g., t test, Welch test).22,23However, this method is often overly optimistic, prone to overfitting, and difficult to reproduce. To overcome these issues, we used integrated algorithms such as mRMR and GBDT to remove redundant features to prevent multicollinearity, and built a predictive model using only high-scoring variables to avoid overfitting. Second, our model allows patients to be evaluated with a single blood draw without the need for additional modalities. This concept is particularly attractive for routine screening of people at high risk of developing disease, such as patients with advanced or severe liver fibrosis, in primary care settings. In these cases where clinically severe liver fibrosis is suspected, it was previously necessary to confirm puncture pathology. However, now only routine serological tests are required to determine the possibility of severe liver fibrosis, so invasive puncture tests can be avoided. Thus, it has obvious advantages in terms of cost and prognosis. Furthermore, our method can be used to build similar model visualizations to distinguish early liver fibrosis from severe liver fibrosis, does not require specially trained clinicians, is practical and convenient for clinicians, and is of great value for clinical promotion.
In this study, we also hoped to improve the diagnostic performance of our model by identifying more specific markers and building a model based on a combination of known serologically associated features. While Zeng et al. used laboratory markers such as B2 macroglobulin, haptoglobin, and apolipoprotein A1 that are not commonly used in most hospitals, we integrated some of the most common serological markers.twenty fourAlthough these laboratory markers may show higher accuracy than conventional serological markers, they are not suitable for real clinical application.Our results show that, among the five conventional serological markers used to build the ML model, HBV-DNA contributed the most to the model, which is consistent with the recommendation of some guidelines that patients with high HBV-DNA levels should be evaluated for non-invasive liver fibrosis.4,25HBV DNA is a marker of viral replication. In chronic HBV infection, disease progression is a dynamic process, and the infection state lasts for a long time. For patients with chronic HBV infection in the indeterminate phase, dynamic follow-up observation is necessary because test results alone may not be able to accurately assess the natural history stage. Studies have shown that HBV DNA levels are correlated with significant fibrosis in HBeAg(-) CHB patients. HBV DNA levels may be able to predict liver fibrosis in HBeAg(-) CHB patients with indications for biopsy.26,27.
In addition, two coagulation factors, including INR and TT, were integrated into the model, although the two coagulation factors are closely related in clinical practice.28,29This may result in overfitting of the model and overestimating the role of clotting factors. However, when we calculated the VIF values of the relevant factors, no collinearity was found. Therefore, we speculate that the contribution of clotting factors to the model should not be overestimated.
It is well known that in many studies it is more difficult to distinguish F0-1 from F2-4.30,31This is because the heterogeneity of liver fibrosis in patients with liver fibrosis F ≥ 2 is more severe than that in patients with liver fibrosis F ≥ 3 and 4, which generally reduces the accuracy of all classification strategies. Indeed, our study results confirm that the accuracy of the DT model in identifying patients with liver fibrosis grade F2 is the lowest (AUC 0.891 in the training cohort and AUC 0.876 in the validation cohort). However, the DT model shows high accuracy and good stability for each fibrosis grade in the two cohorts, especially in identifying cirrhosis (F4), indicating that the model can be used to refine phenotypes in large-scale studies. Our study results also showed that when the model was used to classify risk predictions in the two cohorts or across the entire cohort, the highest overall recognition rate of patients with cirrhosis (F4) was higher than that of patients with other stages of liver fibrosis. These results suggest that our ML model could be part of a more accurate preclinical detection pathway to assess liver cirrhosis and could potentially be used to screen and treat cirrhosis in HBV-infected patients in routine clinical settings, although this needs to be validated in prospective studies.
This study has some limitations. First, this study is a retrospective study, which may lead to the simulation of retrospective statistics that rely on many assumptions. Future studies should focus on the development of prediction and classification models based on prospective studies, which will allow the prediction model to be evaluated, revised, and reevaluated using time-evolution information. Second, the model itself should be further optimized by better engineering and further developed by more comprehensively integrating other clinical data to improve the overall performance of the model and more accurately diagnose liver fibrosis staging noninvasively. Finally, our study did not investigate the performance of the ML model for classifying CHB patients from different ethnic groups, which is also worth further studying in the future. Of course, this study highlights that as a conceptual study, it can provide a certain foundation for actual clinical practice in the future, but this future still has a long way to go.
In conclusion, this study demonstrated that the ML model is more accurate than traditional serological mixture biomarkers in assessing all four stages of liver fibrosis in patients with CHB. Furthermore, the results of this study advance the goal of assessing liver fibrosis in patients with CHB and improving existing prognostic models, thereby facilitating the design and evaluation of future prospective studies, as well as clinical disease monitoring and treatment. We also hope to further refine and extend this study to make this model applicable to a broader range of liver fibrotic diseases.
