An artificial intelligence–based framework for longitudinal risk prediction in patients with cirrhosis through trajectory modeling: A multi-center prospective study in time-series cohorts.

Abstract

e16011 Background: Surveillance for hepatocellular carcinoma (HCC) in adults with cirrhosis relies on fixed-interval AFP testing and imaging examination, which are designed to detect established disease rather than predict its occurrence. Consequently, HCC is often diagnosed between surveillance visits or after progression beyond a clinically actionable stage. Thus, there is a need for approaches that prospectively predict HCC occurrence before clinical detection using routinely available data. Methods: We conducted a multicenter longitudinal study including retrospective model development, independent external validation, and randomized prospective interventional cohorts. Using routinely collected laboratory data, we developed the Temporal Estimation Model for Predicting Occurrence of HCC (TEMPO) to estimate individualized probabilities of HCC occurrence within predefined 6- and 12-month horizons. Model performance was assessed using AUC in development and validation cohorts. In the prospective cohort, participants were randomized 1:1 to TEMPO-guided or AFP-based surveillance (threshold 20 ng/mL). Sensitivity comparisons were performed after calibrating TEMPO to matched diagnostic specificity with AFP. Results: A total of 6,909 participants were included in the development cohort, 6,022 in the external validation cohort, and 4,408 in the prospective cohort. In the development cohort, TEMPO demonstrated strong discrimination for near-term HCC prediction, with AUCs of 0.937 (95% CI 0.923–0.950) and 0.895 (0.876–0.918) for the 6- and 12-month horizons, respectively. In the external validation cohort, corresponding AUCs were 0.921 (0.903–0.939) and 0.861 (0.838–0.885). In the prospective interventional cohort, after calibration to matched diagnostic specificity, sensitivity for incident HCC detection at 6 and 12 months was higher with TEMPO-guided surveillance than with AFP-based surveillance (0.93 and 0.80 vs. 0.59 and 0.50). Among participants with incident HCC, TEMPO-guided surveillance was associated with lower tumor burden at diagnosis, including a higher likelihood of single-tumor presentation (relative proportion 0.78, 0.65–0.94 at 6 months; 0.52, 0.34–0.79 at 12 months), smaller dominant tumor diameter (0.55, 0.41–0.69 at 6 months; 0.61, 0.47–0.77 at 12 months), a higher proportion of BCLC stage 0 disease (0.12, 0.06–0.25 at 6 months; 0.25, 0.10–0.60 at 12 months), and higher use of curative-intent treatments, including surgical resection (0.69, 0.52–0.91 at 6 months; 0.63, 0.42–0.94 at 12 months). Conclusions: TEMPO uses routinely collected baseline laboratory data to predict individualized probabilities of hepatocellular carcinoma occurrence within 6 and 12 months in adults with cirrhosis, enabling prediction of HCC before it becomes clinically detectable.