Standard durations of direct acting antivirals (DAAs; 8–12 weeks) can be a barrier to HCV treatment initiation and completion among marginalised populations. This study developed a machine learning model to predict short-duration (4–6 weeks) DAA response using baseline clinical factors with potential to improve treatment uptake, cost-effectiveness, and health system efficiency.
Baseline data from several short-duration DAA clinical trials and treatment discontinuations from real-world cohort studies were used. Multiple machine learning models were evaluated. Nested cross-validation was employed to optimise model hyperparameters and assess performance. Clinical utility was evaluated using Area Under Receiver Operator Characteristics (AUROC), Area Under Precision Recall Curve (AUPRC) and Matthews Correlation Coefficient (MCC). Threshold optimisation strategies were applied to balance model accuracy and DAA costs. Statistical analyses were conducted to estimate HCV RNA cutoffs predictive of failure.
Of 264 receiving short-duration DAAs (median 42 days; interquartile range 28-42), 94 (36%) experienced treatment failure. Predictors of failure included shorter durations, higher HCV RNA, higher AST–ALT ratio, genotype 3, and DAA class. The Elastic Net (regularised logistic regression) model demonstrated strong performance (AUROC: 83%; AUPRC: 73%). The Youden Index threshold balanced sensitivity (81%) and specificity (76%) with MCC of 0.56. A cost-optimized threshold, prioritizing retreatment minimization, achieved high sensitivity (98%) but reduced specificity (51%). HCV RNA cutoffs predictive of failure were higher for protease+NS5A inhibitors vs. NS5A+NS5B inhibitors.
Predictive models using readily available baseline clinical data can identify individuals likely to respond to short-duration DAAs, with tailored thresholds enhancing clinical utility. Such models, if validated in larger datasets could facilitate HCV elimination efforts by improving treatment uptake, particularly for people who inject drugs, are homeless or incarcerated.
Predicted probabilities of treatment failure are plotted against baseline HCV RNA levels, as estimated by the Elastic Net model, with bootstrapped 95% confidence intervals. The curves illustrate differences in HCV RNA cutoffs for predicting treatment response for NS5A+NS5B and PI+NS5A regimens, stratified by treatment duration and genotype. Results are shown for genotype 3 infections treated with (A) 28-day and (B) 42-day durations; and for non-genotype 3 infections treated (C) 28-day and (D) 42-day durations with. AST-ALT ratio was fixed at the mean value in all scenarios, and HCV RNA cutoff points correspond to probabilities crossing the default 0.5 threshold.
Abbreviations: HCV, hepatitis C virus; GT, genotype; PI, protease inhibitor.