The artificial intelligence-driven model for prediction of hepatocellular carcinoma development in chronic hepatitis B patients: Derivation and validation using 11, 111 patients from Asian and Caucasian cohorts

Abstract

Background and aims: Several risk scores have recently been developed to predict hepatocellular carcinoma (HCC) in patients with chronic hepatitis B (CHB). Our aims were to develop and validate an artificial intelligence (AI)-assisted prediction model for HCC risk, and to compare the utility of our model with previous models. Method: Using Gradient Boosting Machine (GBM), a model was developed from 6, 051 chronic hepatitis B patients on entecavir or tenofovir therapy from 4 tertiary hospitals in Korea. Two external validation cohorts were independently established: Asian cohort (3, 420 patients from 3 Korean centers) and Caucasian PAGE-B cohort (1, 640 from 11 Western centers). The primary outcome was HCC development. The performance of PLAN-B model was compared to previous models using Harrell’s c-index. Results: In the derivation cohort and two validation cohorts, mean ages were 48.6–52.8 years and 60.2–70.6% were male. Cirrhosis was present in 26.9–50.2% at baseline. A total of 595 patients (9.8%) developed HCC during a median 5.2 years of follow-up in the derivation cohort. A model using 10 parameters at baseline was derived and showed a good prediction performance (c-index, 0.79). This model was designated as the PLAN-B model and consisted of the presence of cirrhosis, age, platelet count, antiviral agent (tenofovir or entecavir), baseline serum levels of HBV DNA, alanine aminotransferase, and bilirubin, and HBeAg status, which are listed in order of importance according to Shapley values. PLAN-B showed significantly better predictive power than previous models (PAGE-B, modified PAGE-B, REACH-B, and CU-HCC) in both Asian (c-index, 0.79 vs. 0.64–0.74; all p < 0.001) and Caucasian validation cohorts (c-index, 0.81 vs. 0.57–0.79; all p <0.05 except modified PAGE-B [p = 0.22]). When the patients were grouped into 4 groups using a probability result from GBM algorithm, respective groups showed significantly different risk of HCC development from other groups in both Asian (Fig. 1A) and (Fig. 1B) cohorts and there was good correlation between expected and observed risk (Fig. 1A and 1B). Specifically, the group with the minimal risk group (11.2% of Asian cohort and 8.8% of Caucasian cohort) showed <1% of HCC risk during 10 years of follow-up. Conclusion: This AI-based PLAN-B model provides the best prediction power for HCC risk in Asian as well as Caucasian patients with CHB under entecavir or tenofovir treatment.