Intelligent virtual biopsy can predict fibrosis stage in chronic hepatitis C, combining ultrasonographic and laboratory parameters, with 100% accuracy

Background and aims: The gold standard in evaluating fibrosis stage in chronic hepatitis C (CHC) patients is liver biopsy, a costly and invasive procedure. Alternatively, transient elastography (FibroScan®) performs well in identifying severe fibrosis or cirrhosis, but is less accurate in identifying lower degrees of fibrosis. We recently built a predictive model based on artificial intelligence for staging liver fibrosis in CHC patients, using several non-invasive approaches – routine laboratory tests and basic ultrasonographic parameters – and liver stiffness measurement (LSM). The accuracy of the model was 100%. In this paper, our aim was to investigate if it is possible to reach the same accuracy without LSM. Methods: From more than 100 attributes we first selected the relevant ones by: 1) removing features or cases that do not add any useful information 2) sorting remaining features and assigning ranks based on importance, and 3) identifying the subset of features to use in subsequent AI models by preserving only the most important predictors. For modeling we used decision trees (DT), based on C5.0 algorithm. To improve the DT accuracy, we over-sampled unbalanced categorical variables, and we used boosting an advanced AI technique allowing a team of DT to vote the fibrosis stage. Results: Of the 255 CHC patients, 165 for training and 90 for validation, 5.1% (13/255) had Metavir F0, 33.73% (86/255) F1, 42.75% (109/255) F2, and 18.42% (47/255) F3; Metavir F4 patients were excluded being less than 5%. The relevant features were: cholesterol, caudate lobe diameter, thickness of the abdominal aortic wall, aspartate aminotransferase, preperitoneal fat thickness, splenic vein diameter, time averaged maximum velocity in hepatic artery, time averaged mean velocity in hepatic artery, flow acceleration in hepatic artery, hepatic artery peak systolic velocity. Combining these 10 attributes, the DT were able to predict each fibrosis stage, according to METAVIR scoring system, with 100% accuracy even without LSM. Conclusions: To our best knowledge this, and our previous work, is the first attempt to predict individual Metavir F scores instead of a binary partition of the CHC patients in “significant“ and “nonsignificant“ fibrosis, using a large patients data set –319 cases before preprocessing which removed 64 uninformative cases – and reaching the best published accuracy (100%). This predictive models need to be validated in larger, independent series.