Researchers have developed a machine learning (ML) model to determine first-line antiviral treatment for patients with chronic hepatitis B (CHB) based on the individual risk for hepatocellular carcinoma (HCC), according to a study in the American Journal of Gastroenterology.
The Prediction of Liver cancer using Artificial intelligence-driven model for Network-antiviral Selection for hepatitis B (PLAN-S) was based on data from 13,970 patients with CHB from South Korea, Hong Kong, and Taiwan.
A total of 11,333 treatment-naive patients with CHB initially treated with either entecavir (ETV) or tenofovir disoproxil fumarate (TDF) were enrolled from 16 tertiary hospitals in Korea between 2007 and 2018. A derivation cohort of 6,790 patients (men, 63%) who began antiviral drug use before 2014, a Korean validation cohort of 4,543 participants (men, 58%) who initiated antiviral therapy after 2014, and an independent Hong Kong-Taiwan validation cohort of 2,637 patients (men, 70%) for external validation were included.
Participants for whom an ML-predicted HCC risk at year 5 was higher during ETV vs TDF treatment by 0.25% were the TDF-superior group, and the others were the TDF-nonsuperior group. The primary outcome was confirmed diagnosis of HCC.
After a median follow-up of 77.0 (IQR, 59.0-91.0) months, 764 patients (11.3%) in the derivation cohort had HCC. In the full cohort, the incidence of HCC was significantly lower in the TDF-treated group vs the ETV-treated group (log-rank P <.001).
PLAN-S was created using the random survival forests algorithm and 8 baseline characteristics: sex, cirrhosis, platelet count, antiviral agent (ETV or TDF), hepatitis B e antigen positivity, serum levels of alanine aminotransferase, total bilirubin, and hepatitis B virus (HBV) DNA.
In the derivation cohort, PLAN-S (c-index, 0.78) had superior discriminant function vs previous models (c-index, 0.63-0.74; all P <.001) and similar performance to PLAN-B (c-index, 0.79; P =.18). Comparable results occurred in patients of the Korean validation cohort. For the Hong Kong-Taiwan validation cohort, PLAN-S (c-index, 0.67) had comparable discriminant function except for PLAN-B (c-index, 0.72; P <.001) and modified PAGE-B (c-index, 0.72; P =.004). After classification of participants into 4 risk groups with use of the PRED value derived from PLAN-S, each group had a significantly different HCC risk in all cohorts.
Based on the PREDETV-PREDTDF value at year 5, participants in each cohort were categorized into the TDF-superior (PREDETV-PREDTDF≥0.25%) or TDF-nonsuperior groups (PREDETV-PREDTDF <0.25%). In the derivation, Korean validation, and Hong Kong-Taiwan validation cohorts, 65.3%, 63.5%, and 76.4%, respectively, of patients were in the TDF-superior group.
TDF was associated with a lower cumulative HCC incidence vs ETV in the entire derivation cohort and the TDF-superior group of the derivation cohort (both log-rank P <.001), although no significant difference occurred between the 2 drugs in the TDF-nonsuperior group of the derivation cohort (log-rank P =.44).
In univariable Cox analysis, no significant difference was observed based on type of nucleos(t)ide analogs in the TDF-nonsuperior group (TDF vs ETV: hazard ratio [HR], 1.16; 95% CI, 0.80-1.69; P =.44); and, TDF was associated with a reduced risk for HCC vs ETV in the TDF-superior group (HR, 0.60; 95% CI, 0.49-0.73; P <.001).
Comparable findings were observed in the Korean and the Hong Kong-Taiwan validation cohorts. The cumulative HCC incidence was significantly decreased when participants received TDF in the TDF-superior groups of the Korean (HR, 0.73; 95% CI, 0.54-0.98; P =.03) and Hong Kong-Taiwan validation cohorts (HR, 0.69; 95% CI, 0.49-0.96; P =.03). The HCC risk was comparable between patients who received ETV vs TDF in the TDF-nonsuperior groups of the Korean (HR, 1.22; 95% CI, 0.57-2.61; P =.61) and Hong Kong-Taiwan validation cohorts (HR, 1.29; 95% CI, 0.65-2.55; P =.47).
Among several limitations, only Asian patients with CHB were included, and PLAN-S incorporated only 8 baseline variables. Also, the predictive performance of PLAN-S was relatively low in the Hong Kong-Taiwan cohort, and only HCC development was taken into account when establishing a cutoff value to classify subgroups.
“This ML-based model can offer a personalized antiviral treatment option in patients with CHB using 8 parameters easily available in clinical practice,” the study authors noted.
Disclosure: Some of the study authors declared affiliations with biotech, pharmaceutical, and/or device companies. Please see the original reference for a full list of authors’ disclosures.
This article originally appeared on Gastroenterology Advisor
Hur MH, Park MK, Yip TC-F, et al. Personalized antiviral drug selection in patients with chronic hepatitis B using a machine learning model: a multinational study. Am J Gastroenterol. Published online April 10, 2023. doi:10.14309/ajg.0000000000002234