•Normal tissue complication probability (NTCP) model can guide treatment selection.•NTCP for radiation-induced liver toxicity (RILT) has been developed.•RILT risk varied among different Child–Pugh and viral hepatitis infection status.•Estimated NTCP model and ΔNTCP for RILT are specific to the patient subgroups.•ΔNTCP with uncertainty in individualized subgroups is helpful to select treatment. To predict the probability of radiation-induced liver toxicity (RILT) and implement the normal tissue complication probability (NTCP) model-based approach considering confidence intervals (CIs) to select patients for new treatment techniques, such as proton beam therapy, based on a certain NTCP reduction (ΔNTCP) threshold for primary liver cancer patients. Common Toxicity Criteria for Adverse Events (CTCAE) grade ≥2 RILT was scored. The Lyman NTCP models predicting the probability of CTCAE grade ≥2 RILT as a function of the fraction-size adjusted mean liver dose (MLD), using reference fraction size = 2 Gy/fraction and α/β ratio = 2 Gy, were fitted using the maximum likelihood method. At certain combinations of MLDs, ΔNTCP with a CI was evaluated by the delta method. Of the 239 patients, the incidence of CTCAE grade ≥2 RILT was 55% (46% in the Child–Pugh (CP)-A vs. 81% in the CP-B/C, p < 0.001). Among 180 CP-A patients, 40% who had viral hepatitis infections experienced toxicity vs. 32% in the nonhepatitis subgroup. The MLD was 18 Gy in the toxicity group vs. 16.1 Gy in the nontoxicity group (p = 0.002). The estimated NTCP model parameters specific to the patient subgroups and the ΔNTCP with CI assuming a particular CP classification and viral hepatitis infection status were considerably different which possible changed treatment decision. Patients with CP-A and viral hepatitis infection or CP-B/C cirrhosis had greater susceptibility to CTCAE grade ≥2 RILT. The estimated NTCP and ΔNTCP for individual patients along with a consideration of uncertainties improve the reliability of the NTCP model-based approach.