Prediction of Treatment Response in Hepatitis B and Hepatitis C Coinfected Patients Using a Leakage-Proof, Internally Validated Logistic Regression Pipeline.

Coinfection of hepatitis B virus and hepatitis C virus presents great difficulty in treatment procedures, especially in terms of prediction of the response to the direct-acting antiviral therapy. In view of this, developing adequate prediction models would play a vital role in ensuring better personalization and effectiveness in terms of treatment. The objective of this paper is to develop and validate prediction models for the identification of patients' responses to direct-acting antiviral therapy in chronic HBV/HCV coinfected patients following treatment completion. This retrospective clinical prediction study involved the use of medical records of 154 patients with HBV/HCV coinfection treated with the help of direct-acting antiviral therapy in Holy Family Hospital, Rawalpindi, Pakistan. Sixteen predictors were used to develop logistic regression models to predict treatment responses at 4 and 12 weeks. In addition, to counter the problems of class imbalance, the synthetic minority oversampling technique was applied to the datasets, while nested stratified cross-validation was used for hyperparameter tuning and model validation. Performance was evaluated through different performance metrics such as ROC-AUC, accuracy, precision, recall, and F1-score. The performance of the two models was very good at each endpoint. For the endpoint of 4 weeks, the ROC-AUC score was 0.858 (95% CI 0.746-0.942), the accuracy was 0.760, and the F1-score was 0.851. In the case of the 12-week endpoint, the ROC-AUC score was 0.850 (95% CI 0.769-0.916), the accuracy was 0.786, and the F1-score was 0.856. The important predictors were HCV genotypes, age, body mass index, hemoglobin, and liver function test results. Good model calibration was evident from the calibration graphs, which showed slight deviation from the ideal calibration line at both endpoints. The current study developed prediction models for treatment response in HBV/HCV coinfected patients based on clinical and laboratory information from baseline. The models showed very good internal validity, but the 12-week model performed slightly better than the 4-week model in terms of classification balance. Such models represent useful tools for decision-making in case of a personalized approach.