AI-guided automated identification of patients at high risk for hepatocellular carcinoma in large US and global datasets

BACKGROUND & AIMS: Hepatocellular carcinoma (HCC) screening targets patients with cirrhosis but lacks risk stratification and misses non-cirrhotic cases. Thus, we evaluated whether LIver cancer RIsk Computation (LIRIC) models built from routinely collected electronic health record (EHR) data can predict 3-year HCC risk in general and cirrhosis populations and generalize across diverse US and international cohorts. METHODS: We conducted a multicenter study using longitudinal EHR data from 64 US and 29 ex-US (Latin America/Asia-Pacific) healthcare organizations. Individuals aged ≥40 years at cohort identification and without previous HCC were included; cases were identified using ICD codes 6-36 months prediagnosis. US models were developed in general and cirrhosis cohorts (46,679 cases; 1,128,202 controls), internally externally validated by site, race/ethnicity, and time, and evaluated in a simulated 'silent' deployment. We benchmarked model-derived risk against population incidence using standardized incidence ratios (SIRs). External validation applied US models to ex-US cohorts; region-specific ex-US models were also trained. Logistic regression and neural networks (a liver risk computation logistic regression model and a liver risk computation neural network model) were compared using AUC, calibration, and geometric mean of overestimation (GMOE) risk ratios. RESULTS: In the US general population, LIRICNN achieved AUC 0.93 (95% CI 0.9218-0.9289) using 46 features. Internal-external AUCs averaged 0.93 across sites and race/ethnicity strata, with good calibration (GMOE 0.89; 95% CI 0.862-0.911). At a screening threshold corresponding to SIR ≈ 31, sensitivity was 48% and specificity was 97%. US-trained models applied to ex-US data yielded lower AUCs (0.84), but region-specific retraining restored performance (AUC 0.94). CONCLUSIONS: LIRIC models accurately stratified HCC risk using routine EHR data in both general and cirrhosis populations and could be adapted for international settings via local retraining. These results support LIRIC as a scalable foundation for risk-based HCC surveillance strategies and future prospective implementation studies. IMPACT AND IMPLICATIONS: Current HCC screening targets patients with cirrhosis but lacks risk stratification and misses non-cirrhotic cases; therefore, we evaluated whether LIRIC models built from routinely collected EHR data can predict 3-year HCC risk in general and cirrhosis populations and generalize across diverse US and international cohorts. These findings are important for hepatologists, primary care clinicians, and health systems because they show that routine EHR data can support accurate and scalable HCC risk stratification across diverse populations. In practice, LIRIC could serve as a foundation for risk-based HCC surveillance strategies and future prospective implementation studies using data already collected in routine care. Given that this was a retrospective modeling study and international performance required local retraining, these findings should be interpreted as supporting future implementation rather than immediate universal adoption.