Liver fibrosis is characterized by the accumulation of an extracellular matrix (ECM), which distorts the hepatic architecture. Liver fibrosis progression could commonly be found in chronic hepatitis B patients after an extensive period of time during which liver biochemical indices are found to be predominately or even persistently abnormal. Due to the limitation of liver biopsy, noninvasive evaluation of liver fibrosis is thus of great clinical interests in order to assess the risk for liver fibrosis dynamically, or identify and monitor the patients who should be considered antiviral or other types of therapy.
Age, platelet, AST, ALT, GGT, etc., as routine biochemical markers, had been well known predictors of significant liver fibrosis[14, 29, 30]. In the present study, based on common biochemical parameters including routine and serum markers, we constructed a three-layer neural network which extended a back-propagation learning algorithm by introducing probabilistic treatment of the Bayesian inference technique for the synaptic weight. Results of sensitivity analysis showed the importance of various predictors. As could be seen, the most important variables influencing the prediction of significant liver fibrosis were age, AST, platelet and GGT. These results were consistent with some of the earlier studies[14, 16]. Sensitivity analysis showed that the neural network using Bayesian approaches could achieve its predictive purpose.
As demonstrated by ROC curves, the predictive accuracy of the artificial neural network was reasonably high in the training, validation, and testing sets. The AUROC were 0.883, 0.884 and 0.920, respectively. In view with some noninvasive parameters directly or indirectly related to fibrogenesis, the most prevalent are the Fibrotest and the Actitest for necro-inflammator activity. They are based on GGT, TBIL, haptoglobin, α2-macroglobulin, apolipoprotein A1, and for the Actitest additionally on ALT. Previous studies on Fibrotest and Actitest have been validated with the ranges of the AUROC of 0.75- 0.89 in CHB patients[14, 25, 32, 34]. Although the two indices could provide better predictive values according to different criteria, they were calculated with a patented and complicated algorithm, and it was difficult for physicians to use them to identify the states of liver fibrosis.
Hui AY constructed and validated a multivariate logistic regression model using body mass index, platelet, Alb, and TBIL level to predict advanced fibrosis, and the AUROC were 0.765-0.803. Zeng MD also constructed a scoring system with forward logistic regression, which was expressed by the following formula: -13.995 + 3.22l g(α2-macroglobulin)+ 3.096 lg(age) + 2.254 lg(GGT) + 2.437 lg(HA). The AUROC of this model in the training and validation groups were 0.84 and 0.77, respectively. The logistic regression, a generalized linear model used for binomial regression, is used for prediction of the probability of occurrence of an event by fitting data to a logistic curve. The artificial neural networks, a non-linear statistical data modeling tool, can be used to model complex relationships between inputs and outputs. Therefore, the present study provided evidence that the three-layer neural network model based on routine and serum markers was superior to other indices or models for identification of individual with or without significant fibrosis.
In our study, there were 126 patients with significant liver fibrosis, which only accounted for 27.7% of 455 CHB patients with liver biopsy. In other words, 329 (72.3%) patients without significant liver fibrosis had undergone liver biopsy, and they could bear the damages from such an invasive procedure. From a more practical point of view, we wanted to reduce the number of liver biopsy procedures and also identify all CHB patients with significant fibrosis. Therefore, our study evaluated the influence of the different cutoff points on the accuracy of ROC. When we chose a high cutoff point, the number of CHB patients at high risk for significant liver fibrosis was few, and also fewer patients needed further liver biopsy or other examinations. But, there was a low sensitivity and many CHB patients with significant liver fibrosis could be missed. The purpose of predicting the state of liver fibrosis is to identify CHB patients with significant liver fibrosis or at high risk for liver fibrogenesis to prevent them from further liver fibrogenesis. Thus, we should choose a lower cutoff point to improve the sensitivity, and to reduce the number of missed CHB patients at high risk for liver fibrogenesis. So, we considered a probability value of 0.33 as a cutoff value. CHB patients with a probability value > 0.33 were considered with significant liver fibrosis or at high risk for liver fibrogenesis. In our study, all CHB patients with significant liver fibrosis would be identified. 47.4% (55/116) of the CHB patients would be free of liver biopsy and also wouldn't be missed.