Lesion selection and external evaluation of clinical-imaging survival models after transarterial chemoembolization for hepatocellular carcinoma

Background: Pretreatment computed tomography (CT) describes tumor characteristics in patients receiving transarterial chemoembolization (TACE) for hepatocellular carcinoma (HCC), but whether measuring more tumors improves mortality predictions in a new cohort remains unclear. We assessed the added value of different lesion-selection scopes beyond shared clinical predictors. Methods: We compared a model using five clinical variables with models adding CT measurements from the largest, two largest, or all retained annotated tumors. All models used the same regularized discrete-time survival estimator. Evaluation used repeated nested cross-validation in 228 WAW-TACE patients and frozen external application in 69 HCC-TACE-Seg patients, with 15 deaths by day 365. We assessed first-year probability error, discrimination, and calibration, followed by local recalibration using small external samples. Results: Internally, adding CT-derived lesion features did not establish an improvement in prediction accuracy. Externally, the all-lesion model had the highest discrimination point estimates, but its advantage over the clinical-only model was uncertain. Its mortality probabilities were less accurate than those of the clinical-only model, with calibration estimates consistent with overly extreme predicted risks. Local recalibration reduced its median held-out prediction error at day 365, but the median adjusted error remained higher than that of the adjusted clinical model. Simple tumor-burden predictors also did not improve external probability accuracy over the clinical baseline. Conclusions: In these cohorts, measurements from more annotated tumors did not yield more reliable first-year mortality probabilities. Comparisons of clinical-imaging models should evaluate external probability accuracy and calibration alongside patient ranking.