(P-627) Evaluation of Machine-learning Algorithms for Predicting Direct-acting Antiviral Treatment Failure in Patients with Chronic Hepatitis C: An HCV-TARGET Analysis

Background: Despite low failure rates, the number of patients with hepatitis C virus (HCV) who failed to achieve a sustained virologic response (SVR) after treatment is increasing as the result of widespread adoption of direct acting antivirals (DAAs).

Objectives: To develop and validate machine-learning algorithms to predict DAA treatment failure among patients with HCV.

Methods: We used HCV-TARGET registry data, a multicenter, prospective, observational study to identify HCV patients ≥18 years who initiated all-oral DAA regimens before December 2018 and who had a virologic outcome. We randomly divided patients into training and validation samples. Treatment failure was defined as not achieving SVR (undetectable plasma RNA level at least 12 weeks after completing DAA treatment). Potential predictors (n=365) included socio-demographics, clinical factors, adverse events, and virologic data. We applied multivariable logistic regression (MLR), elastic net (EN), random forests (RF), gradient boosting machine (GBM), and feedforward neural networks (FNN) to predict treatment failure. We assessed multiple prediction performance measures (e.g., C-statistic, number needed to evaluate [NNE] to identify one treatment failure).

Results: Patients in the training and validation samples had similar characteristics (mean age of 57 years, 60% male, 66% white, and 36% cirrhosis). Of 6525 HCV patients, 6217 patients (95.3%) achieved SVR while 308 patients (4.7%) experienced DAA failure. The four machine-learning approaches performed similarly in predicting DAA failure including EN (C-statistic=0.74, 95%CI=0.69-0.79), FNN (C-statistic=0.75, 95%CI=0.70-0.80), RF (C-statistic=0.74, 95%CI=0.69-0.80), and GBM (C-statistic=0.72, 95%CI=0.67-0.78). All four approaches outperformed MLR (C-statistic=0.51, 95%CI=0.46-0.57) with EN using the least number of predictors (n=32). Using the Youden Index risk score threshold, EN had a 58.4% sensitivity, 77.8% specificity, and NNE of 14. Over 60% of patients with DAA failure were classified by EN in the top three highest risk deciles (7 to 22 DAA failures per 100 patients). The important predictors identified by EN included male sex, advanced liver diseases (i.e., decompensated cirrhosis, hepatocellular carcinoma), albumin ( <3.5g/dL) and total bilirubin (>1.2 g/dL) levels, alcohol or tobacco use, short DAA treatment duration ( < 8 weeks), and use of proton pump inhibitors or immunosuppressants.

Conclusions: Machine-learning algorithms appear to perform effectively for risk prediction and stratification of DAA failure.