Validation of a machine learning approach using FIB-4 and APRI scores assessed by the metavir scoring system: A cohort study.

BACKGROUND AND STUDY AIM: The study aim was to improve and validate the accuracy of the fibrosis-4 (FIB-4) and aspartate aminotransferase-to-platelet ratio index (APRI) scores for use in a potential machine-learning (ML) method that accurately predicts the extent of liver fibrosis. PATIENTS AND METHODS: This retrospective multicenter study included 69,106 patients with chronic hepatitis C planned for antiviral therapy from January 2010-December 2014 with liver biopsy results. FIB-4 and APRI scores were calculated and their performance for predicting significant liver fibrosis (F3-F4) assessed against the Metavir scoring system. ML was used for feature selection and reduction to identify the most relevant attributes (CfsSubseteval/best first) for prediction. RESULTS: In this study, 57,492 (83.2%) patients were F0-F2, and 11,615 (16.8%) patients were F3-F4. The revalidation of FIB-4 and APRI showed lower accuracy and higher disagreement with the biopsy results, with AUCs of 0.68 and 0.58, respectively. FIB-4 diagnosed fewer (14%) F3-F4 patients, and the high specificity and negative predictive values of FIB-4 and APRI reflected the low prevalence of F3-F4 in the study population. Out of 15 attributes, age (>35 years), AFP (>6.5 ng/ml), and platelet count (<150,000/mm3) were the most relevant risk attributes, and patients with one or more of these risk factors were likely to be F3-F4, with a classification accuracy of ≤ 92% and receiver operating characteristics area of 0.74. CONCLUSION: FIB-4 and APRI scores were not very accurate and missed diagnosing most of the F3-F4 patients. ML implementation improved medical decisions and minimized the required clinical data to three risk factors.

Acoustic radiation forced impulse-based splenic prediction model using data mining for the noninvasive prediction of esophageal varices in hepatitis C virus advanced fibrosis.

BACKGROUND: Esophageal varices (EV) are serious complications of hepatitis C virus (HCV) cirrhosis. Endoscopic screening is expensive, invasive, and uncomfortable. Accordingly, noninvasive methods are mandatory to avoid unnecessary endoscopy. Acoustic radiation forced impulse (ARFI) imaging using point shear wave elastography as demonstrated with virtual touch quantification is a possible noninvasive EV predictor. We aimed to validate the reliability of liver stiffness (LS) and spleen stiffness (SS) by an ARFI-based study together with other noninvasive parameters for EV prediction in HCV patients. Also, we aimed to evaluate the diagnostic performance of a new simple prediction model (incorporating SS) using data mining analysis. PATIENTS AND METHODS: This cross-sectional study included 200 HCV patients with advanced fibrosis. Labs, endoscopic, ultrasonographic, LS, and SS data were collected. Their accuracy in diagnosing EV was assessed and a data mining analysis was carried out. RESULTS: Ninety patients (22/46% of F3/F4 patients) had EV (39/30/18/3 patients had grade I/II/III/IV, respectively). LS and SS by ARFI showed high significance in differentiating not only patients with/without EV (P = 0.000 for both) but also correlated with the grading of varices (R = 0.31 and 0.45, respectively; P = 0.000 for both). Spleen longitudinal diameter (SD), splenic vein diameter (SVD), platelets to spleen diameter ratio, LOK index, and FIB-4 score were the best ultrasonographic and biochemical predictors for the prediction of EV [area under receiver operating characteristic (AUROC) 0.79, 0.76, 0.76, 0.74, and 0.71, respectively]. SS (using ARFI) had better diagnostic performance than LS for the prediction of EV (AUROC = 0.76 and 0.70, respectively). The diagnostic performance increased using data mining to construct a simple prediction model: high probability for EV if [(SD cm) × 0.17 + (SVD mm) × 0.06 + (SS) × 0.97] more than 6.35 with AUROC 0.85. CONCLUSION: SS by ARFI represents a reliable noninvasive tool for the prediction of EV in HCV patients, especially when incorporated into a new data mining-based prediction model.

Data Mining and Machine Learning Algorithms Using IL28B Genotype and Biochemical Markers Best Predicted Advanced Liver Fibrosis in Chronic Hepatitis C.

IL28B single nucleotide polymorphism (rs12979860) is an etiology-independent predictor of hepatitis C virus (HCV)-related hepatic fibrosis. Data mining is a method of predictive analysis which can explore tremendous volumes of information from health records to discover hidden patterns and relationships. The current study aims to evaluate and compare the prediction accuracy of scoring system like aspartate aminotransferase-to-platelet ratio index (APRI) and fibrosis-4 (FIB-4) index versus data mining for the prediction of HCV-related advanced fibrosis. This retrospective study included 427 patients with chronic hepatitis C. We used data mining analysis to construct a decision tree by reduced error (REP) technique, followed by Auto-WEKA tool to select the best classifier out of 39 algorithms to predict advanced fibrosis. APRI and FIB-4 had sensitivity-specificity parameters of 0.523-0.831 and 0.415-0.917, respectively. REPTree algorithm was able to predict advanced fibrosis with sensitivity of 0.749, specificity of 0.729, and receiver operating characteristic (ROC) area of 0.796. Out of the 16 attributes, IL28B genotype was selected by the REPTree as the best predictor for advanced fibrosis. Using Auto-WEKA, the multilayer perceptron (MLP) neural model was selected as the best predictive algorithm with sensitivity of 0.825, specificity of 0.811, and ROC area of 0.880. Thus, MLP is better than APRI, FIB-4, and REPTree for predicting advanced fibrosis for patients with chronic hepatitis C.

Predictors of Virological Response in 3,235 Chronic HCV Egyptian Patients Treated with Peginterferon Alpha-2a Compared with Peginterferon Alpha-2b Using Statistical Methods and Data Mining Techniques.

Despite the appearance of new oral antiviral drugs, pegylated interferon (PEG-IFN)/RBV may remain the standard of care therapy for some time, and several viral and host factors are reported to be correlated with therapeutic effects. This study aimed to reveal the independent variables associated with failure of sustained virological response (SVR) to PEG-IFN alpha-2a versus PEG-IFN alpha-2b in treatment of naive chronic hepatitis C virus (HCV) Egyptian patients using both statistical methods and data mining techniques. This retrospective cohort study included 3,235 chronic hepatitis C patients enrolled in a large Egyptian medical center: 1,728 patients had been treated with PEG-IFN alpha-2a plus ribavirin (RBV) and 1,507 patients with PEG-IFN alpha-2b plus RBV between 2007 and 2011. Both multivariate analysis and Reduced Error Pruning Tree (REPTree)-based model were used to reveal the independent variables associated with treatment response. In both treatment types, alpha-fetoprotein (AFP) >10 ng/mL and HCV viremia >600 × 10(3) IU/mL were the independent baseline variables associated with failure of SVR, while male gender, decreased hemoglobin, and thyroid-stimulating hormone were the independent variables associated with good response (P < 0.05). Using REPTree-based model showed that low AFP was the factor of initial split (best predictor) of response for either PEG-IFN alpha-2a or PEG-IFN alpha-2b (cutoff value 8.53, 4.89 ng/mL, AUROC = 0.68 and 0.61, P = 0.05). Serum AFP >10 ng/mL and viral load >600 × 10(3) IU/mL are variables associated with failure of response in both treatment types. REPTree-based model could be used to assess predictors of response.

Application of data mining techniques to explore predictors of HCC in Egyptian patients with HCV-related chronic liver disease.

BACKGROUND: Hepatocellular carcinoma (HCC) is the second most common malignancy in Egypt. Data mining is a method of predictive analysis which can explore tremendous volumes of information to discover hidden patterns and relationships. Our aim here was to develop a non-invasive algorithm for prediction of HCC. Such an algorithm should be economical, reliable, easy to apply and acceptable by domain experts. METHODS: This cross-sectional study enrolled 315 patients with hepatitis C virus (HCV) related chronic liver disease (CLD); 135 HCC, 116 cirrhotic patients without HCC and 64 patients with chronic hepatitis C. Using data mining analysis, we constructed a decision tree learning algorithm to predict HCC. RESULTS: The decision tree algorithm was able to predict HCC with recall (sensitivity) of 83.5% and precession (specificity) of 83.3% using only routine data. The correctly classified instances were 259 (82.2%), and the incorrectly classified instances were 56 (17.8%). Out of 29 attributes, serum alpha fetoprotein (AFP), with an optimal cutoff value of ≥50.3 ng/ml was selected as the best predictor of HCC. To a lesser extent, male sex, presence of cirrhosis, AST>64U/L, and ascites were variables associated with HCC. CONCLUSION: Data mining analysis allows discovery of hidden patterns and enables the development of models to predict HCC, utilizing routine data as an alternative to CT and liver biopsy. This study has highlighted a new cutoff for AFP (≥50.3 ng/ml). Presence of a score of >2 risk variables (out of 5) can successfully predict HCC with a sensitivity of 96% and specificity of 82%.