Premium
Machine learning algorithms for predicting direct‐acting antiviral treatment failure in chronic hepatitis C : An HCV‐TARGET analysis
Author(s) -
Park Haesuk,
LoCiganic WeiHsuan,
Huang James,
Wu Yonghui,
Henry Linda,
Peter Joy,
Sulkowski Mark,
Nelson David R.
Publication year - 2022
Publication title -
hepatology
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 5.488
H-Index - 361
eISSN - 1527-3350
pISSN - 0270-9139
DOI - 10.1002/hep.32347
Subject(s) - medicine , algorithm , logistic regression , random forest , machine learning , hepatitis c virus , hepatitis c , viral load , artificial intelligence , immunology , human immunodeficiency virus (hiv) , virus , computer science
Background and Aims We aimed to develop and validate machine learning algorithms to predict direct‐acting antiviral (DAA) treatment failure among patients with HCV infection. Approach and Results We used HCV‐TARGET registry data to identify HCV‐infected adults receiving all‐oral DAA treatment and having virologic outcome. Potential pretreatment predictors ( n = 179) included sociodemographic, clinical characteristics, and virologic data. We applied multivariable logistic regression as well as elastic net, random forest, gradient boosting machine (GBM), and feedforward neural network machine learning algorithms to predict DAA treatment failure. Training ( n = 4894) and validation ( n = 1631) patient samples had similar sociodemographic and clinical characteristics (mean age, 57 years; 60% male; 66% White; 36% with cirrhosis). Of 6525 HCV‐infected adults, 95.3% achieved sustained virologic response, whereas 4.7% experienced DAA treatment failure. In the validation sample, machine learning approaches performed similarly in predicting DAA treatment failure (C statistic [95% CI]: GBM, 0.69 [0.64–0.74]; random forest, 0.68 [0.63–0.73]; feedforward neural network, 0.66 [0.60–0.71]; elastic net, 0.64 [0.59–0.70]), and all four outperformed multivariable logistic regression (0.51 [0.46–0.57]). Using the Youden index to identify the balanced risk score threshold, GBM had 66.2% sensitivity and 65.1% specificity, and 12 individuals were needed to evaluate to identify 1 DAA treatment failure. Over 55% of patients with treatment failure were classified by the GBM in the top three risk decile subgroups (positive predictive value: 6%–14%). The top 10 GBM‐identified predictors included albumin, liver enzymes (aspartate aminotransferase, alkaline phosphatase), total bilirubin levels, sex, HCV viral loads, sodium level, HCC, platelet levels, and tobacco use. Conclusions Machine learning algorithms performed effectively for risk prediction and stratification of DAA treatment failure.