Machine Learning Predicts Hepatocellular Carcinoma Risk from Routine Clinical Data: A Large Population-Based Multicentric Study

Machine-learning models trained on routine multimodal clinical data accurately predict future hepatocellular carcinoma risk across large population cohorts, providing an interpretable framework for precision screening.

Abstract

Hepatocellular carcinoma (HCC) is a highly fatal tumor, for which risk stratification is crucial yet remains challenging. In this study, we develop an interpretable machine learning (ML) framework for HCC risk stratification based on routinely collected clinical data. We utilize prospectively collected multimodal data from more than 900,000 individuals and 983 cases of HCC across two population-scale cohorts: the UK Biobank study (development) and the All of Us Research Program (external testing). We assess individual and cumulative contributions of data modalities, including demographics, lifestyle, health records, blood, genomics, and metabolomics. Our final random forest–based models significantly outperform all publicly available state-of-the-art risk scores on both internal and external test sets. We demonstrate robustness across ethnic subgroups, provide comprehensive interpretability, and release all code, model weights, and a web calculator for external validation and agentic integration. Our study presents PRE-Screen-HCC, a robust and interpretable ML framework for HCC risk stratification and early detection.

Significance:

Using data from population-scale cohorts, we develop and externally validate an ML framework for HCC risk stratification. Models trained on routine clinical data outperform published scores, perform on par with metabolomics and genomics, generalize across subgroups, and remain interpretable.

See related commentary by Foda, p. 1252

Introduction

Hepatocellular carcinoma (HCC) is the fifth most common malignancy and the third leading cause of cancer-associated death. Due to its increasing incidence, HCC is considered a major public health concern, especially in structurally disadvantaged regions, both locally and globally (1–6).

Current screening protocols for HCC rely largely on previously diagnosed liver cirrhosis as the primary inclusion criterion, as the majority of HCC cases arise in patients with cirrhosis (7, 8). However, this strategy overlooks the clinical challenges associated with diagnosing cirrhosis itself, as fibrosis often progresses asymptomatically (9). This leads to a substantial number of HCC cases being diagnosed only at late disease stages, drastically worsening prognosis (7, 8, 10). Additionally, the strategy fails to consider multifactorial risk factors such as lifestyle, preexisting health conditions, blood tests, or omics signatures, which could identify a broader at-risk population and more accurately represent the individual risk for HCC (11–14). This gap is particularly critical as the prevalence of metabolic dysfunction–associated steatotic liver disease (MASLD) and its related HCC cases continues to increase, and screening for HCC is resource-intensive and requires skilled clinicians (15–18). This highlights the need for a screening strategy that is more affordable, more inclusive, and more efficient at detecting patients with a high risk of HCC (19). Imaging-based screening could then be specifically targeted at this group.

Integrating the vast array of patient data to assess disease risk represents a significant challenge in modern medicine, particularly as screening is typically conducted by general practitioners who must possess a broad knowledge base across a wide range of diseases and risk factors. Machine learning (ML) algorithms present a promising solution to this integration challenge and could further reduce healthcare disparities if developed cautiously (20–22). Several studies have shown the superiority of ML algorithms over traditional regression analyses or few parameter-based decision support systems as implemented in current guidelines (1, 9, 23–34). However, a major criticism is the lack of generalizability due to small cohorts, lack of published model weights, a lack of external testing (35), and the usage of highly curated, retrospective datasets that do not necessarily represent real-world scenarios (21).

In response to these challenges, we developed a novel, noninvasive multimodal prescreening tool that integrates diverse risk factors into a comprehensive risk stratification model. Our study uses the extensive data available from the UK Biobank (UKB) and the All of Us Research Program (AOU), both of which include extensive questionnaires, electronic health records (EHR), blood parameters, and genomics for more than 500,000 participants, with ongoing follow-up since 2006 for UKB (36–39) and extensive retrospective documentation for AOU (40, 41). By reducing reliance on the diagnosis of cirrhosis before HCC and instead including a wide array of noninvasive, affordable, and available risk factors, we aim to improve the early detection of HCC. This approach not only promises to expand the efficacy of current screening practices but also extends its reach to under-resourced regions, potentially transforming HCC prognosis on a global scale.

Our study specifically investigates the individual, cumulative, and potentially synergistic contributions of various data modalities, such as genotypic information versus lifestyle and blood data, to highlight the distinct advantages and augmented value each contributes to HCC risk stratification. We hereby especially emphasize the current underutilization of more accessible and cost-effective data types (Fig. 1A–C) and show that the integration of multimodal data into ML models outperforms previous approaches for risk stratification of HCC development.

Figure 1.

Study concept. A, The task of predicting HCC occurrence was divided into prediction from a healthy cohort (“All”) and prediction among patients at risk (“PAR”). Multimodal data from UKB were extracted, and scenarios were set up according to the availability of data on a patient’s trajectory in the healthcare system. B, ML architecture with an inner-layer fivefold cross-validation, using a grouped-split approach, in which each split (indicated by small squares) combines 4–5 assessment centers together, with each split serving four times as a training set and once as a validation set. Training data were solely generated from UKB centers within England, as indicated on the schematic map of Great Britain. The final model is a mean vote (M) built from the five generated ML models. M was then applied to the withheld test set built from UKB centers in Scotland/Wales (+Newcastle, for an 80:20 balance between train and test) with numerical prediction output. C, Evaluation and independent external testing of the model, including classification, time-to-event analysis, and subgroup analysis. Thresholds are applied to rule in for screening (when > “high” threshold) or to rule out for screening (when < “low” threshold).

Results

HCC Cases in UKB Represent the General Population

We hypothesized that population-based cohorts are well suited to build HCC prediction models. First, they have superior generalizability to the general population compared with cohorts from tertiary centers, and second, they provide rich data on phenotypic features even years before cancer diagnosis (36, 39). A total of 538 eligible HCC cases were observed in UKB (Supplementary Fig. S1A; Fig. 2A) and 445 HCC cases in AOU (Supplementary Fig. S1B). The mean time to HCC diagnosis in UKB was 8.7 years, with a mean age at diagnosis of 70 ± 6.8 years and 62.1 ± 10.2 years in AOU (Supplementary Fig. S2A and S2B). This aligns with previous observations that report substantial differences in age at onset for HCC across countries (42). Incidence was comparable with the overall HCC incidence in the UK with 6 to 10/100,000 new cases per year (Fig. 2A; Table 1; ref. 43), whereas it was more frequent in the AOU cohort (15–30/100,000; Table 2). All 22 centers of data acquisition showed a similar prevalence of HCC (Fig. 2B; Supplementary Table S1). Of 538 HCC cases, 399 had a positive cancer record in the respective national cancer registries, corroborating binary outcomes as a robust and accurate endpoint. We manually selected demographic and lifestyle features (n = 51) based on the extensively characterized risk factors of HCC (alcohol consumption, smoking, social status; Table 1; Supplementary Table S2; refs. 44, 45). We then performed a phenome-wide association study (PheWAS) of HCC in UKB, matched for age, sex, and body mass index (BMI) to select broad disease-level features (Supplementary Table S3). Our PheWAS revealed a multitude of significant correlations (Fig. 2C; Supplementary Tables S3–S7), expectedly associating mostly with liver diseases but also, for example, alcoholism and diabetes (Fig. 2D). Notably, only 31% of patients with HCC had received an EHR-coded diagnosis of cirrhosis, viral hepatitis, or other chronic liver diseases (CLD) prior to HCC diagnosis. Forty-one percent of patients received a diagnosis of CLD in the 2 years after the first HCC diagnosis, and 28% never received a diagnosis of CLD (Fig. 2E; Supplementary Tables S8 and S9). This underlines the high fraction of underdiagnosis of CLD before the diagnosis of HCC, as most HCCs develop in cirrhotic livers (7). It further challenged the development of time-based models due to inconsistent event recording, in which the time component was less reliable than the binary information of whether HCC occurred. To account for different pretest probabilities of HCC, we defined two cohorts for which all modeling steps were executed. First, the entire UKB cohort (All) and, second, the “patients-at-risk” population (PAR), which included all patients with previous CLD (steatotic liver diseases, viral hepatitis, or cirrhosis) and cases with elevated liver enzymes at baseline examination [International Classification of Diseases (ICD) codes in Supplementary Table S4; for thresholds, see “Methods”]. We included all blood count and serum parameters available at baseline, confirming 33 features that were significantly associated with the diagnosis of HCC after controlling for age, sex, BMI, and CLD (Fig. 2F; Supplementary Tables S10 and S11). Common single-nucleotide polymorphisms (SNP) known to increase the risk of liver diseases, such as MASLD, cirrhosis, and HCC, were included after confirmation via linear regression analysis on UKB data (Fig. 2G; Supplementary Table S12; refs. 12, 46–53). Nuclear magnetic resonance (NMR) spectroscopy metabolomics on 248,266 participants was investigated for possible correlating effects after correction for age, sex, and BMI (Supplementary Fig. S2C; Supplementary Table S13). Of 143 metabolomic features, we observed 109 significant associations with HCC (Supplementary Table S14).

Figure 2.

UKB encompasses representable risk factors of HCC. A, Histogram of first occurrences of HCC diagnosis in EHR in UKB. Cases discarded for analysis are shown in dark gray (for exclusion criteria, see “Methods”). B, Distribution of HCC cases in UKB centers on a map of Great Britain, displayed by country association and mapping to training or validation cohort, mapped as the number of cases per 500,000. C, Manhattan plot of phenome-wide associations for individuals with HCC (C22.0) vs. propensity-matched control population [on age, sex, BMI; with P values (−log₁₀)]. Highlighted are associations with P values ≤ 0.05 (corrected for multiple testing by Bonferroni correction). D, Prevalence of the most common disease codes in EHR or self-reported for the HCC cohort, patients-at-risk cohort, and control group, sorted by highest prevalence in the HCC group. E, Ranking of etiologies of HCC cases (n = 538) with total numbers and percentages. Bars indicate time points, with “first EHR” being the first-ever NHS-documented hospital stay, “Prior to HCC” the last stay before HCC diagnosis and “0–2 years after HCC” indicating the first diagnosis in the first 2 years after HCC diagnosis. No LD, no liver disease. F, Associations of blood biomarkers with HCC in PAR, displayed as odds ratios (OR) and 95% confidence intervals for a 1-SD increase in each biomarker on the natural log scale, adjusted for age, sex, and BMI. Significance is defined as false discovery rate (FDR) controlled P < 0.01 in a linear model, marked with * and full saturation. Error bars are limited to Y limits for better readability. For mapping to UKB field IDs, see Supplementary Table S5. G, Associations of SNPs with HCC occurrence are displayed as log₂-transformed ORs + 95% confidence interval, split between heterozygous (het) and homozygous (hom) occurrence, with color indicating the direction of effect (blue = reduced risk, red = increased risk), and significance (Bonferroni-corrected P < 0.05) indicated by opacity: transparent = not significant, opaque = significant.

Table 1.

Baseline characteristics of the UKB study population stratified by sex and HCC status.

Characteristic	Female					Male
	Overall	HCC	No HCC	P valueb	q valuec	Overall	HCC	No HCC	P valueb	q valuec
	N = 273,245a	n = 130a	n = 273,115a			N = 229,028a	n = 408a	n = 228,620a
Age (years)	56.4 (±8)	61 (±6.1)	56.3 (±8)	<0.001	<0.001	56.7 (±8.2)	61.4 (±6.1)	56.7 (±8.2)	<0.001	<0.001
BMI	27.1 (±5.2)	29 (±6.5)	27.1 (±5.2)	0.001	0.016	27.8 (±4.2)	30.3 (±5.1)	27.8 (±4.2)	<0.001	<0.001
Waist circumference (cm)	84.8 (±12.5)	90.5 (±16.2)	84.8 (±12.5)	<0.001	0.001	96.9 (±11.3)	104.4 (±13.3)	96.9 (±11.3)	<0.001	<0.001
Weight (kg)	71.5 (±14.1)	75.6 (±17.3)	71.5 (±14.1)	0.007	0.10	85.9 (±14.3)	91.8 (±16.9)	85.9 (±14.3)	<0.001	<0.001
Standing height (cm)	162.5 (±6.3)	161.6 (±6.7)	162.5 (±6.3)	0.13	>0.9	175.6 (±6.9)	174 (±7)	175.6 (±6.9)	<0.001	<0.001
Ethnicity				0.2	>0.9				0.5	>0.9
Unknown	1,266 (0.5%)	<5	1,266 (0.5%)			1,512 (0.7%)	5 (1.2%)	1,507 (0.7%)
Caucasian	257,320 (94%)	123 (95%)	257,197 (94%)			215,172 (94%)	385 (94%)	214,787 (94%)
Mixed	1,849 (0.7%)	<5	1,849 (0.7%)			1,104 (0.5%)	<5	1,103 (0.5%)
Asian or Asian British	4,580 (1.7%)	<5	4,578 (1.7%)			5,292 (2.3%)	10 (2.5%)	5,282 (2.3%)
Black or Black British	4,649 (1.7%)	<5	4,648 (1.7%)			3,406 (1.5%)	<5	3,404 (1.5%)
Chinese	988 (0.4%)	<5	988 (0.4%)			583 (0.3%)	<5	582 (0.3%)
Other	2,593 (0.9%)	<5	2,589 (0.9%)			1,959 (0.9%)	<5	1,955 (0.9%)
Multiple deprivation index	17 (±13.7)	18.9 (±13.7)	17 (±13.7)	0.10	>0.9	17.5 (±14.2)	21 (±16.1)	17.5 (±14.2)	<0.001	<0.001
Systolic blood pressure (mm Hg)	135.3 (±18.7)	138.5 (±19.7)	135.3 (±18.7)	0.071	>0.9	140.6 (±17.1)	144.7 (±17.7)	140.6 (±17.1)	<0.001	<0.001
Medication				<0.001	<0.001				<0.001	<0.001
No medication	188,802 (69%)	68 (52%)	188,734 (69%)			153,704 (67%)	167 (41%)	153,537 (67%)
Metabolic	62,863 (23%)	56 (43%)	62,807 (23%)			75,324 (33%)	241 (59%)	75,083 (33%)
Hormones	21,580 (7.9%)	6 (4.6%)	21,574 (7.9%)			<5	<5	<5
Diabetes mellitus	9,585 (3.5%)	27 (21%)	9,558 (3.5%)	<0.001	<0.001	14,711 (6.4%)	132 (32%)	14,579 (6.4%)	<0.001	<0.001
Family diabetes	52,853 (19%)	20 (15%)	52,833 (19%)	0.3	>0.9	41,613 (18%)	88 (22%)	41,525 (18%)	0.086	>0.9
Pack-years	6.5 (±11.8)	7.8 (±11.8)	6.5 (±11.8)	0.2	>0.9	10.4 (±16.7)	20.9 (±23.5)	10.4 (±16.7)	<0.001	<0.001
Alcohol (g/day)	6.6 (±8.7)	4.7 (±8.9)	6.6 (±8.7)	0.018	0.3	14.5 (±15.2)	17.6 (±21.2)	14.5 (±15.1)	0.003	0.045

Baseline characteristics of study participants (N = 502,309) stratified by sex and HCC status. Data are presented as mean (±SD) for continuous variables and n (%) for categorical variables. P values were calculated using the Student t test for continuous variables and χ² test for categorical variables. Statistical significance was defined at α = 0.05, values below this threshold are displayed in bold. All continuous variables were assessed at time of UKB enrollment. q values represent Bonferroni-adjusted P values to account for multiple testing.

^aMean (±SD); n (%).

^bWelch two-sample t test; Pearson’s χ² test.

^cBonferroni correction for multiple testing.

Table 2.

Baseline characteristics of AOU study population stratified by sex and HCC status.

Characteristic	Female					Male
	Overall	HCC	No HCC	P valueb	q valuec	Overall	HCC	No HCC	P valueb	q valuec
	N = 76,720a	n = 140a	n = 76,580a			N = 46,583a	n = 189a	n = 46,394a
Age (years)	54.1 (±16.2)	60.9 (±12.5)	54.1 (±16.2)	<0.001	<0.001	59.3 (±15.5)	64.7 (±8.4)	59.3 (±15.5)	<0.001	<0.001
BMI	30.8 (±8.2)	30 (±7.2)	30.8 (±8.2)	0.2	>0.9	29.7 (±6.6)	28.9 (±5.3)	29.7 (±6.6)	0.031	0.4
Waist circumference (cm)	95.8 (±17.3)	97 (±15.2)	95.8 (±17.3)	0.3	>0.9	102.2 (±15.5)	102.3 (±12.4)	102.2 (±15.5)	0.9	>0.9
Weight (kg)	81.3 (±22.8)	78.2 (±21.1)	81.3 (±22.8)	0.088	>0.9	91.9 (±22)	87 (±18)	91.9 (±22)	<0.001	0.003
Standing height (cm)	162.3 (±7.1)	161.2 (±7.1)	162.3 (±7.1)	0.070	>0.9	175.7 (±7.8)	173.3 (±7.5)	175.7 (±7.8)	<0.001	<0.001
Self-reported ethnicity				0.086	>0.9				0.10	>0.9
Asian	1,718 (2.2%)	<20	1,716 (2.2%)			1,073 (2.3%)	<20	1,069 (2.3%)
Black/African American	13,308 (17%)	33 (24%)	13,275 (17%)			6,750 (14%)	36 (19%)	6,714 (14%)
Latinx	14,073 (18%)	31 (22%)	14,042 (18%)			6,128 (13%)	35 (19%)	6,093 (13%)
Middle Eastern	320 (0.4%)	<20	319 (0.4%)			310 (0.7%)	<20	309 (0.7%)
More than one	1,139 (1.5%)	<20	1,139 (1.5%)			514 (1.1%)	<20	512 (1.1%)
No answer	2,069 (2.7%)	<20	2,062 (2.7%)			1,503 (3.2%)	<20	1,494 (3.2%)
Pacific Islander	58 (<0.1%)	<20	58 (<0.1%)			58 (0.1%)	<20	58 (0.1%)
White	44,035 (57%)	66 (47%)	43,969 (57%)			30,247 (65%)	102 (54%)	30,145 (65%)
Multiple Deprivation Index	0.3 (±0.1)	0.3 (±0.1)	0.3 (±0.1)	<0.001	0.005	0.3 (±0.1)	0.3 (±0.1)	0.3 (±0.1)	<0.001	<0.001
Systolic blood pressure (mm Hg)	125 (±17.4)	129.3 (±19.1)	125 (±17.4)	0.008	0.10	130.2 (±16.6)	130.9 (±16.8)	130.2 (±16.6)	0.6	>0.9
Medication				0.9	>0.9				0.3	>0.9
Hormones	831 (1.1%)	<20	830 (1.1%)			313 (0.7%)	<20	311 (0.7%)
Metabolic	15,187 (20%)	29 (21%)	15,158 (20%)			11,642 (25%)	38 (20%)	11,604 (25%)
No medication	60,702 (79%)	110 (79%)	60,592 (79%)			34,628 (74%)	149 (79%)	34,479 (74%)
Diabetes mellitus	21,471 (28%)	78 (56%)	21,393 (28%)	<0.001	<0.001	15,918 (34%)	102 (54%)	15,816 (34%)	<0.001	<0.001
Family diabetes	10,635 (14%)	<20	10,623 (14%)	0.091	>0.9	4,259 (9.1%)	<20	4,247 (9.2%)	0.2	>0.9
Pack-years	4.8 (±11.7)	8 (±17.5)	4.8 (±11.7)	0.037	0.5	8.8 (±17.2)	10.4 (±17.1)	8.8 (±17.2)	0.2	>0.9
Alcohol (g/day)	3.4 (±7.2)	1 (±3)	3.4 (±7.2)	<0.001	<0.001	6.2 (±11.6)	5.5 (±14)	6.2 (±11.6)	0.5	>0.9

Baseline characteristics of study participants (N = 123,303) stratified by sex and HCC status. Participants with multiple visits and status change from control group to HCC group over time were only considered for the positive class (most recent entry), resulting in 329 unique HCC patients and 382 patient–timepoint observations. Participants with HCC diagnosis before visit were excluded from the analysis. Data are presented as mean (±SD) for continuous variables and n (%) for categorical variables. P values were calculated using the Student t test for continuous variables and χ² test for categorical variables. Statistical significance was defined at α = 0.05, values below this threshold are displayed in bold. Features with less than 20 individual data points grouped together are masked (displayed as <20) as is the requirement of the AOU. Q values represent Bonferroni-adjusted P values to account for multiple testing.

^aMean (±SD); n (%).

^bWelch two-sample t test; Pearson’s χ² test.

^cBonferroni correction for multiple testing.

Stepwise Study Architecture Mimics Data Availability in Real-World Settings

We investigated the predictive capacities of five different clinically relevant data modalities, namely demographics, EHRs, bloodwork, genomics, and metabolomics, in two ways: first, individually for each modality and, second, using a stepwise approach reflecting the clinical availability of the modalities (Fig. 1A). Predictive capacities were then assessed for both cohorts, All and PAR, resulting in a total of 20 model variations. Training and iterative hyperparameter tuning were performed in a fivefold cross-validation on data from England-based centers of UKB (18 of 22 centers). The final models were then tested on untouched data from the four remaining centers, located either in Scotland or Wales (+ Newcastle; n = 123 of 538 HCC cases in UKB; Fig. 1B; Supplementary Tables S1 and S15), as well as on the entire AOU cohort (Supplementary Table S16).

Compared with more complex deep learning techniques, decision tree–based ML models like random forest classifiers (RFC) have the advantage of interpretability due to feature importance metrics while being more data-efficient than neural networks or transformers (54, 55). In addition to RFC, we included the architectures extreme gradient boosting (XGB), Categorical Boosting (CatBoost), and a simple neural network (multilayer perceptron) in initial benchmarking experiments (Supplementary Fig. S3A and S3B). All models performed similarly; however, RFC performance was most consistent, especially for models trained on smaller cohort sizes (PAR). As RFC is also dominant in literature reports for HCC risk prediction (23, 29), we selected the RFC architecture as our baseline. A grid search for hyperparameter optimization on the internal cross-validation set identified an optimal, slim architecture (max-depth = 3, n_estimators = 50). Subsequently, exclusively this configuration was used to train the final model to further avoid overfitting (see the “Methods” section).

ML Is Superior to Linear Risk Scores for Prediction of HCC Occurrence in the General Population

We hypothesized that the accuracy of ML models would be superior to established liver-related HCC risk assessment scores. We performed a comparative analysis among available scores in the literature: aMAP ß; ref. 56), FIB-4 (57), APRI (58), and NFS (59) scores for the prediction of HCC (Supplementary Fig. S3C and S3D). Herein, the aMAP score achieved the best performance of all linear scores, with an area under the receiver operating characteristic curve (AUROC) of 0.79 (“All”) and 0.83 (“PAR”). We therefore used the aMAP score as our primary benchmark throughout model performance evaluation. We first studied each modality independently (demographics, diagnosis, blood, genomics, and metabolomics, Supplementary Fig. S3E–S3G) and, second, incrementally based on clinical availability (Fig. 3A–G; Supplementary Tables S17 and S18). Evaluation on the withheld UKB test set (90 HCC cases/93,533 controls for “All” and 72 cases/20,958 controls for “PAR”) for independent modalities revealed blood parameters as the most relevant modality (AUROC, 0.86 and 0.87 for All/PAR, respectively), followed by demographics (0.80/0.78), metabolomics (0.79/0.82), diagnosis (0.74/0.73), and SNPs (0.62/0.65).

Figure 3.

Prediction metrics. A, Split violin plots displaying the distribution of predicted probabilities for HCC-negative (left half violin) and HCC-positive (right half violin) participants per trained modality for all patients in the UKB test set. Quartiles (25th, 50th, 75th) are displayed with dotted lines. Participants with liver cirrhosis are separately illustrated with grey circles in absolute numbers. B and C, Model performance on the test set as receiver operating characteristic (ROC) curves [+ area under the curve (AUROC)], as well as PRCs + indicated area under the curve for either all patients (B and C) or patients at risk (E and F) for incremental (additive) models. Each line represents the performance of one mean vote model from fivefold cross-validation. The benchmark score “aMAP” and random guess are depicted by dotted lines. D–F, As in A–C, respectively, but for the PAR subgroup. G, Legend for the contribution of modalities per incremental model. H and I, ROC curves and PRC for models with stepwise reduction of model complexity, as indicated in J, which serves as a legend for the process of reducing the number of features included in models from model C to TOP 15 based on individual feature importance. PRCs for models A–E, as well as aMAP score, with a magnified section for precision = 0–0.2 or 0–0.4. J–L, As in H and I, but for the PAR subcohort.

Clinical routine is multimodal, and a single modality likely never captures the true complexity. We therefore hypothesized (i) that the combination of modalities could improve model performance substantially, and (ii) that, based on previous results, less affordable omics methods might not increase model performance substantially. Incremental models indeed revealed overall superior performance compared with separate models, with a plateau at an AUROC of 0.88 with model C, a combination of demographics, diagnosis, and blood data (Fig. 3A–G). Performance could only be mildly improved by adding genomics and metabolomics information.

Precision–recall curves (PRC), highly relevant for imbalanced data, revealed extremely low performance for all literature benchmarks (Fig. 3C and F), with areas under the PRCs (AUPRC) of 0.00 to 0.02 for “All” and 0.02 to 0.03 for “PAR.” The only exception for this was the prediction of HCC based solely on the ICD code for liver cirrhosis (AUPRC of 0.11 for “All” and “PAR”); however, AUROCs for cirrhosis were low (AUROC, 0.57/0.59 for UKB and 0.72/0.72 for AOU). Meanwhile, our ML models reached AUPRC values of 0.07, 0.06, and 0.11 (All, models C, D, and E) and 0.09, 0.10, and 0.22 (PAR, models C, D, and E). Prediction improvement for models D and E however occurred predominantly for patients at very high risk. Meanwhile, at recall levels relevant for routine screening, all curves aligned, suggesting no additional contribution to overall discriminatory capabilities in a broader population.

Explainability Methods Highlight Known Risk Factors of HCC

We next performed ablation studies to reduce the number of input features for the model (Fig. 3H–L). We employed a two-step approach, focusing first on routinely assessed features and second on the removal of low-relevance features (60). A team of clinicians curated a set of 75 exclusively routine clinical features. For these, we then gradually removed features with the lowest feature importance (TOP 75 features, TOP 30, TOP 15). For “All,” a significant reduction in performance was observed between TOP 75 and TOP 30 (P = 0.0012, two-sided DeLong test), whereas the performance of TOP 30 and TOP 15 did not differ significantly (P = 1, two-sided DeLong test) and all ML models significantly outperformed the literature benchmark scores (Supplementary Table S19). Mirroring this, a performance reduction was observed in AUPRCs from TOP 75 to TOP 30, but not from TOP 30 to TOP 15 (Fig. 3I and L; Supplementary Fig. S3H). This process was repeated with a TOP 30 and a TOP 15 model for the PAR cohort (Fig. 3K and L). Alongside this, we explored whether the model’s strong performance was due to the modeling architecture or the included features by training an RFC exclusively on the variables included in the aMAP score. The resulting aMAP-RFC model modestly, yet significantly, outperformed the original linear aMAP score, confirming that nonlinear modeling can extract additional signal from these few variables (Supplementary Table S19). However, the substantially higher performance of our feature-selected models indicates that predictive gain primarily arises from the choice and combination of features. Further, secondary analyses revealed that models trained exclusively on participants with complete data for all TOP 15 features showed comparable performance to those trained on partially imputed data, demonstrating the robustness of our imputation strategy (Supplementary Fig. S4A–S4D). Moreover, alpha-fetoprotein at baseline alone was a weak predictor of future HCC in the UKB subset with available proteomic data (Supplementary Fig. S5A–S5D, n = 50,940 individuals, n = 61 HCC).

We hypothesized that features relevant to the ML classifiers would reflect the well-characterized risk landscape represented in linear risk scores. We therefore investigated the contributions of individual features to the models’ decisions and compared those across features and modalities. We consistently observed the highest modality importance for blood-based features such as gamma-glutamyltransferase (γ-GT), aspartate aminotransferase (AST), alanine aminotransferase (ALT), and platelets. Importantly, however, feature importance was distributed more nuanced over all features, also incorporating measurements such as insulin-like growth factor 1 or waist circumference. In total, more than 20 features each contributed with >1% (Fig. 4A–D; Supplementary Fig. S6A–S6D; Supplementary Table S20). We further observed substantially different contributions to model prediction per modality. Blood data consistently had the highest contribution, both when assessed as the mean and sum of feature importance per modality. This was followed by metabolomics, demography and lifestyle, EHR, and lastly genomics, which had negligible relevance for the model (Supplementary Fig. S6; Supplementary Table S20).

Figure 4.

Interpretability of HCC risk prediction models. A, Feature importance of the top 15 features as well as the average by feature group (see Supplementary Material for corresponding features and groups) per indicated model, as mean ± SD of the accumulation of the impurity decrease within each tree for cohort “All” and model TOP 15. B, Split violin plot with displayed thresholding as low (green), middle (yellow), and high (red) risk groups, displaying the distribution of predicted probabilities for controls (left half of the violin) and HCC cases (right half of the violin) per trained modality in the test set for cohort “All” and model TOP 15. Quartiles (25th, 50th, 75th) are displayed with dotted lines. C, As in A, but for the PAR cohort. D, As in B, but for the PAR cohort. E and F, Confusion matrices corresponding to prediction scores for model TOP 15 from B and D, respectively, with indicated thresholds. NPV, negative predictive value; PPV, positive predictive value. Coloring indicates the relative percentage per class. G and H, Time-dependent AUROC and AUPRC for TOP 15 All and TOP 15 PAR, with an indication of all patients who developed HCC up until the indicated timeframe. I and J, Kaplan–Meier curves for relative events per risk group over time. * indicates time frames that were removed from the analysis due to the removal of events close to the baseline examination and due to the cutoff of January 1, 2024.

Actionable Risk Scores and Classes for HCC Risk Estimation

To make the continuous risk scores clinically actionable, we implemented two complementary approaches. First, we calibrated the models to ensure predicted probabilities accurately reflect true risk rather than broadly distributed scores between 0 and 1. This calibration was performed using cross-validation on the training data and validated on the withheld UKB test set (Supplementary Figs. S7A–S7E and S8A–S8E). Second, as clinicians need clear decision thresholds rather than continuous nuance, we established a three-class system with a higher “rule-in” and a lower “rule-out” threshold, based on the distribution of prediction values in the fivefold cross-validation. We evaluated these thresholds on the withheld UKB test set, both for UKB “All” (Fig. 4B) and “PAR” (Fig. 4D). In both cohorts, >70% of HCC cases could hereby be classified in the high-risk group, with a positive predictive value of 0.8% (“All”) and 3.3% (“PAR”) and negative predictive values of 0.9999 (“All”) and 0.9995 (“PAR”). High- and medium-risk groups together accounted for 87.8% (“All”) and 87.5% (“PAR”) of HCC cases (Fig. 4E and F). Still, as a low-incidence disease, this performance was challenged by a high number of false positive (FP) cases. We conducted a comprehensive comparison between the true positive (TP) and false negative (FN) groups, as well as the true negative and FP groups. There was no significant difference between the time to cancer for TP (9.6 ± 3.7 years) and FN (9 ± 3.8 years). We further found negligible performance change over time (Fig. 4G and H) and early divergence in risk per year (Fig. 4I and J) despite a lower absolute risk of HCC per year in UKB compared with the literature (43). Notably, we found a significantly higher incidence of FN among females compared with males (Supplementary Table S21). Further emerging patterns in the FN group included a lower prevalence of obesity, smoking, and alcohol consumption.

Generalizable Models for HCC Risk Stratification

Finally, we assessed our models’ generalizability to independent populations, testing them on data from AOU (tier 7) with at least one instance of available blood data and baseline demographic information (n = 123,303 and 445 cases of HCC). Incidence of HCC (Fig. 5A) was higher than in UKB; however, it was comparable with literature reports (61). We observed a clear tendency toward better EHR representation among HCC cases, especially for liver cirrhosis (Fig. 5B; Table 2; Supplementary Table S22). The ethnic distribution among controls and cases was more diverse than in UKB (Fig. 5C; Table 2). We constructed pseudoprospective cohorts by aggregating baseline data for any patient-year with available data and, consecutively, within discrete 5-year windows (pre-2010, 2010–2015, and 2015–2020). For each window, predictions were made prospectively relative to the baseline, evaluating whether HCC occurred within a predefined follow-up horizon of 5 years after the most recently available lab value entry (Supplementary Table S23). Using this setup, we applied the TOP 30, TOP 15, and AMAP-RFC models alongside the linear risk scores aMAP, FIB-4, APRI, and NFS to all eligible AOU participants, as well as the TOP 15-PAR model for all patients who met the inclusion criteria for PAR. All ML models achieved state-of-the-art AUROCs and AUPRCs, comparable with the UKB test cohorts (Fig. 5; Supplementary Tables S24 and S25), indicating generalizability across cohorts and arguing against overfitting as the underlying explanation. Again, ML models outperformed linear risk scores (Fig. 5D and E; Supplementary Fig. S9A and S9B; Supplementary Table S26). Although the relative distribution of prediction values for controls and cases was very similar to UKB, the prediction scores for both All and PAR were consistently higher in AOU compared with UKB: 0.40 ± 0.16 (AOU) versus 0.33 ± 0.13 (UKB) for “TOP 15-All,” P < 0.0001; 0.67 ± 0.19 (AOU) versus 0.62 ± 0.22 (UKB) for HCC cases, P = 0.24; 0.43 ± 0.16 (AOU) versus 0.31 ± 0.14 (UKB) for PAR, P < 0.0001; and 0.66 ± 0.17 (AOU) versus 0.59 ± 0.19 (UKB) for HCC cases, P = 0.03 (for model TOP 15-PAR, see Supplementary Table S27; Supplementary Fig. S9C–S9F). Most likely, this reflects the overall difference in disease burden and method of data collection between the two populations (Supplementary Table S28) and indicates the necessity of calibration and adapted thresholding for distinct populations.

Figure 5.

ML models are generalizable to independent populations. A, Histogram of first occurrences of HCC diagnosis in EHR in All of Us. B, Prevalence of the most common disease codes for the HCC group (red) and control group (gray), sorted by highest prevalence in the HCC group. AKI, acute kidney injury; Arterial HT, arterial hypertension; CKD, chronic kidney disease; Dis., disease; DM, diabetes mellitus; HSM, hepatosplenomegaly. C, Ethnic distribution of all participants (outer circle) and HCC cases (inner circle) in All of Us, corresponding to self-reported survey categories. D and E, Model performance on the AOU test set as receiver operating characteristic (ROC) curves [+ area under the curve (AUC); D], as well as PRCs (E) + indicated area under the curve. Each line represents the performance of one mean vote model from fivefold cross-validation. Benchmark scores are indicated in gray, previous diagnosis of cirrhosis in blue, and random guess is depicted by dark gray dotted lines. F and G, Confusion matrices corresponding to prediction scores from D–E, split by ethnicity (White vs. non-White). H–K, As in D–G but for the patients at risk subcohort. L and M, Shapley feature importance analysis for model TOP 15 All, with each data point representing a single participant. Alk. Phos, alkaline phosphatase; Art., arterial; Diabetes mell., diabetes mellitus; Hep., hepatitis; MCV, mean corpuscular volume; Waist circ., waist circumference. The distance from 0 on the x-axis indicates feature importance; color indicates the direction of the feature.

Again, model performance was better for male than for female cases although with less pronounced differences (Supplementary Table S21; Supplementary Fig. S9G–S9J). Interestingly, model performance for (self-reported) White versus non-White ethnicity populations did not differ significantly and did especially not lean toward better performance for the White population (Fig. 5F and G), despite training on 94% data from “Caucasian” ethnicity, as is the distribution in UKB (Table 1). For PAR, models achieved similar performance (Fig. 5H–K). Feature importance patterns in AOU mirrored those observed in UKB (Fig. 5L and M), and numbers needed to treat were similar (Supplementary Fig. S10A–S10D). Finally, model calibration developed on UKB data (Supplementary Figs. S7 and S8) showed preserved calibration performance when deployed on AOU data (Supplementary Figs. S11A–S11E and S12A–S12E), supporting the models’ clinical utility in diverse healthcare settings.

Discussion

Our study shows that interpretable ML algorithms can accurately stratify individual risk of developing HCC on a population scale. Evaluation on two of the largest available cohorts worldwide shows the superior performance of our developed ML models (Supplementary Table S29) compared with publicly available risk scores and offers clinically meaningful interpretations.

Current HCC screening, with some exceptions (31), centers on cirrhosis, a highly specific but insensitive risk factor for HCC, as the sole inclusion criterion for apparative screening (7, 8, 62). However, compensated cirrhosis can be occult and therefore not diagnosed in a large fraction of patients. Furthermore, the percentage of HCC in noncirrhotic patients is steadily growing, which widens the population at risk (16, 17, 63). This gap in screening facilitates late-stage diagnosis and high mortality of HCC. Noninvasive risk scores such as FIB-4, developed for high-risk patients are nevertheless commonly used as surrogates for liver fibrosis and predictors of liver-related mortality in the general population despite poor performance at the population level (64, 65). We hypothesized that population-based cohorts with longitudinal follow-up are the ideal foundation for generalizable risk stratification models in HCC. These cohorts collect broad baseline data years before clinical endpoints are documented, which enables the simulation of prospective studies. In line with the “Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis” (TRIPOD) and TRIPOD AI, we conducted independent validation using two large and distinct datasets. First, we developed a nonrandom split sample (TRIPOD 2b), in which the UKB data were split by geographic location [England for development; Scotland and Wales (and Newcastle) for validation]. This design introduces nonrandom variation between the development and validation samples, providing a stronger test of model performance than a random split. Second, we performed an entirely independent external validation (TRIPOD 3) utilizing the “AOU” (66, 67).

To account for distinct pretest probabilities between primary and tertiary care centers, we evaluated models in two cohorts: the general population and a subset of patients at risk. Across published scores, ROC-based metrics remained high, potentially due to the correct classification of abundant negatives in the rare-event setting; however, precision and recall were uniformly low (68–70). By contrast, our developed ML models displayed a 3 to 10 times higher precision than literature scores. This improvement is crucial, as, for example, the improvement in AUPRC from 0.01 to 0.04 leads to a reduction in FPs from 32,000 to ∼8,000 (TOP 15 All, UKB), while maintaining recall (68). Presence of cirrhosis prior to HCC remained a highly specific and therefore good “rule-in” discriminator but insufficient as a “rule-out” discriminator of HCC in the general population. Nevertheless, it functions as such in various guidelines (8, 13).

Average prediction values differed substantially across the tested cohorts. We therefore deliberately chose not to emphasize the thresholds developed during fivefold cross-validation on UKB. The variation in prediction values likely reflects well-characterized differences in disease burden, which is lower for UKB and higher for AOU, as AOU predominantly encompasses hospital data (20, 39, 40, 71). Most likely, thresholds will have to be developed independently for certain populations, taking into account the respective disease burden but also economic considerations across populations and healthcare systems (72). Notably, after Platt scaling on UKB, calibration remained good when transported to AOU, with an acceptable calibration slope/intercept and concordant calibration curves. Importantly, in screening, the number needed to screen (NNS) is inversely proportional to detected positives. Aiming for a detection rate of two out of three HCC cases during prescreening (reflecting our rule-in threshold of 0.55) would lead to an NNS of 70 to 80 in the general population or an NNS of 26 in the PAR population (Supplementary Tables S19 and S21). In contrast, “cirrhosis-only” screening programs could have detected only one in five HCC cases even in the unlikely scenario of perfect screening attendance and perfect detection rates (Fig. 2D). Current clinical routine demands an incidence of 1.5% in the screening group for HCC screening to become economically feasible (73) and a minimal annual incidence of 0.2% in the respective population to allow for effective screening (8, 13). Patients in our model’s high-risk group clearly surpass the annual incidence of 0.2% in both cohorts. This remained true even despite a lower general incidence of HCC in the UKB than in the general population and despite our strict inclusion criteria, as we removed HCC cases near baseline and cases not confirmed by cancer registries. These clinically driven data cleaning strategies further corroborate our choice to frame the task as a classification problem rather than a survival problem. The performance of survival models such as random survival forests would be strongly affected by the noisy and often biased longitudinal follow-up in UKB (e.g., reversed causality in EHR coding and inconsistent timing of diagnoses; ref. 74). By contrast, we could reliably confirm binary outcomes through linkage with national cancer registries, enabling clinically meaningful risk stratification.

Our study presents the first systematic comparison of independent data modalities for HCC prediction. We show that training models exclusively on routine clinical data, including standard blood values like transaminases and platelets, together with EHR and lifestyle data, is sufficient for accurate risk stratification for HCC. This is in line both with previous work for other cancer types and with linear risk scores for HCC and liver disease (56, 58, 75–77). Interestingly, genomic data did not increase model performance despite the representable prevalence of known genetic risk factors like PNPLA3 mutations (Fig. 2G; ref. 12). Consistent with previous reports, this questions the relevance of polygenic risk scores for risk prediction of HCC compared with phenotypic modeling (14, 78, 79). Meanwhile, metabolomics did increase model performance to some extent. However, as they are not routinely available, we deliberately proceeded to ablate and validate exclusively the models developed with routine clinical data and satisfactory performance.

Interpretability analyses confirmed liver enzymes, platelets, and demographic features such as age and weight but also arterial hypertension and liver cirrhosis as predictive features, clearly aligning with clinical expertise and literature findings (9, 52, 56, 58, 75, 80–82). Still, certain features contributed less than expected, indicating present biases and calling for cautious interpretation. The social desirability bias could explain the systematic underestimation of the relevance of alcohol consumption in ML models (83). Furthermore, training on populations with better representation of liver cirrhosis as an ICD code might have assigned higher contributions to liver cirrhosis.

Our study has several limitations. First, our models were not developed on biopsy-proven or fibroscan-matched steatosis (9, 84). Second, our analysis, like all population-based studies, relies on accurate encoding of EHR. Notably, evaluating our models exclusively on cases with HCC encoding in both National Health Service (NHS) and national cancer registries, as opposed to NHS-only, revealed a reduced number of FNs and FPs compared with evaluation on all EHR-encoded cases. Still, a fraction of approximately 2% to 13% of HCC cases were wrongly assigned to the low-risk group, for UKB and AOU test sets alike. Analysis of these misclassified cases revealed that our classifiers demonstrated a higher tendency to overlook HCC in female patients compared with male patients. This could be explained by an overall less prominent HCC phenotype in females or the higher prevalence in males, challenging the assumption that a combined model for all sexes is actually feasible. Interestingly, this performance gap was less pronounced in the AOU cohort, with furthermore no differences in performance for the “White” ethnicity versus other ethnicities. This suggests good generalizability, which is crucial as SLD prevalence varies across populations (22, 71). However, validation in other populations, especially Asian populations with a higher prevalence of viral hepatitis, is still pending. This currently limits claims of generalizability to other etiologies of HCC.

In clinical routine, estimation of annual risk bears greater clinical relevance compared with all-time risk of HCC. We therefore validated PRE-Screen-HCC pseudoprospectively on the 5-year risk of HCC in the All of Us cohort, consistently outperforming linear risk scores and unexpectedly not dropping in predictive accuracy over time. Nonetheless, prospective clinical trials will be needed before clinical adoption can be recommended. Given the low incidence of HCC, developing clinical trials dedicated solely to HCC screening will be challenging. A more feasible approach may be to set up multipurpose trials that screen for liver morbidity in general and HCC alike, such as the LiverAIM trial or the Structured Early detection of Asymptomatic Liver Cirrhosis (SEAL) study introduced for routine primary care checkups (85, 86). As in most cases, steatotic liver diseases are only diagnosed after the diagnosis of HCC; such multipurpose trials could help elucidate whether linking HCC screening to prior diagnosis of liver cirrhosis is still appropriate.

From a technical perspective, our models can be integrated effortlessly for further research purposes and, once a medical device, into clinical practice. Our TOP 15 random forest–based models for the All and PAR cohorts are each built from 15 affordable and routinely available parameters. Users can validate or deploy these models in three complementary ways: first, through an interactive web calculator for single-patient inference (https://huggingface.co/spaces/schneiderlab/PRE_SCREEN_HCC); second, via the accompanying Python package for local batch processing and integration into existing pipelines; and third, within agentic workflows (87, 88) through command-line interfaces or via the model context protocol (https://modelcontextprotocol.io/). Finally, long-term integration should ideally happen interoperably with hospital information systems, in which input data can be automatically retrieved for a multitude of models, as ML predictions enter clinical routine in virtually every discipline of healthcare (89–91).

Our study reveals critical gaps in the current screening process for HCC. It addresses a clinically relevant question in an interpretable way with actionable output. PRE-Screen-HCC outperforms established linear risk calculators for HCC by leveraging interpretable ML and extensive population-based data. We envision PRE-Screen-HCC, given appropriate validation, serving as a first-line, low-cost prescreening tool during routine clinical check-ups to identify individuals at elevated risk of HCC. These patients may benefit from consecutive apparative screening despite not meeting guideline-defined screening criteria. This could pave the way for personalized, molecular prevention and risk assessment, with significant implications for earlier intervention and potentially curative treatment.

Methods

Study Objective

This study is a predictive analytic study with the objective of creating new digital biomarkers that estimate the risk of future HCC development. It includes individual-level data from the data sources UKB from the UK and from AOU from the United States. Data overlap between cohorts can be excluded. All study items were executed adhering to the TRIPOD (66), TRIPOD AI (67), and ESMO GROW (92) reporting standards.

Ethics Disclaimer

All participants in this study were enrolled through research programs with formal ethical oversight and documented informed consent procedures. Participants in the UKB provided written informed consent at assessment centers before any data collection or linkage, and the UKB operates under a formal ethics and governance framework with ongoing research ethics committee approval. AOU protocols and participant materials, including written informed consent, were reviewed and approved by the AOU single Institutional Review Board in compliance with US human subjects research regulations. All research in the present study was conducted in accordance with recognized ethical guidelines, including the Declaration of Helsinki, and relevant national regulations governing human subjects research. Our study does not include confidential information. All research procedures were conducted exclusively on anonymized patient data.

We acknowledge the ethical complexities of categorizing individuals in medical research, especially when categorizing by ethnicity. Although any form of categorization risks perpetuating discrimination, the complete omission of such considerations in medical AI could paradoxically reinforce healthcare disparities. ML models are sensitive to training data composition and can silently perpetuate or amplify existing biases if their performance across different populations remains untested. By explicitly examining model performance across ethnic groups, we aim to ensure equitable predictive accuracy and identify potential disparities in model generalizability that require attention. This approach aligns with the broader goal of developing inclusive healthcare solutions while remaining mindful of the need to handle demographic data with appropriate sensitivity and scientific rigor. Our categorizations follow standardized reporting guidelines while recognizing that such classifications are social constructs and cannot fully capture human diversity.

UKB Cohort

UKB (RRID:SCR_012815) is an expansive prospective cohort study initiated between 2006 and 2010. The UKB enrolled a diverse population of individuals aged 37 to 73 years, encompassing a total of 502,411 participants, with 502,309 participants currently available due to the removal of consent from 102 participants. Initial evaluations were comprehensive, incorporating an array of anthropometric measurements, biospecimen collection, and the administration of multiple structured questionnaires. Ethical approval for the UKB study was granted by the North West Multi-centre Research Ethics Committee. Participants in the UKB provided written informed consent at assessment centers before any data collection or linkage. Comprehensive methodologic details, data accessibility, and acquisition protocols are publicly disclosed on the UKB’s official website (http://www.ukbiobank.ac.uk). Participants were actively encouraged to partake in an inaugural clinical assessment, which was subsequently followed by longitudinal monitoring. All enrollees provided informed consent for genotypic analyses and the longitudinal linkage of their data to medical records. Periodic follow-up evaluations have been instituted to monitor alterations in health status and lifestyle variables.

Primary Survey Data

Baseline characteristics were obtained through structured questionnaires. These self-reported measures included lifestyle factors such as alcohol consumption and smoking history, as well as medical history encompassing liver disease, diabetes, and obesity. Estimated values (alcohol consumption per day; pack-years) were limited to the 99.9th percentile to minimize the influence of extreme outliers while preserving the overall distribution of the data. Physiologically plausible limits were set to further curate the dataset (Supplementary Tables S2 and S16). Ethnicities in the “AOU” were inferred from the questionnaires “self-reported race” and “self-reported ethnicity.”

Electronic Health Records

EHR derived from the NHS and self-reported diagnoses were merged with EHR through manual ICD mapping (Supplementary Table S7). Patients with an EHR entry of HCC before or in the first year after their first UKB visit were excluded from the analysis (n = 36; Fig. 2A) to prevent training on data from undetected HCC cases. Records were supplemented with UK death registers, with C22.0 documented here being treated equally to encoding via EHR. Training was performed on this data, whereas evaluation was restricted to all patients whose diagnosis of HCC was confirmed by data from the UK National Cancer Registry (UKB Resource 115558; Fig. 2B). EHR for a subset of 136 ICD codes (Supplementary Tables S4 and S8; Fig. 2C and D) were included based on positive associations identified in the PheWAS (Fig. 2E). These comprised a set of CLDs, signs of portal hypertension, cardiovascular risk factors, and gastrointestinal cancers except hepatobiliary cancers (Fig. 2E). To prevent leakage of future diagnoses to the training data while ensuring a rich representation of EHR codes in the input training data, diagnosis codes up until 5 years after the initial UKB visit were collected as features, truncated at least 1 year before the first diagnosis of HCC. The EHR of “All of Us” is curated from a variety of sources and then harmonized using the Observational Medical Outcomes Partnership (OMOP) Common Data Model to be stored in data dictionaries. Specific data relevant to our study were then extracted on their workbench platform using built-in cohort and dataset builders. Diagnostic instances of all relevant ICD codes for each individual were tabulated along with shifted dates, as any one individual may have the same diagnosis code recorded in their health record multiple times from multiple medical visits.

Blood Count and Biochemistry

Blood assays in UKB (category 100080) were obtained at baseline assessment, with biological processing explained in detail on the UKB’s official website. Extracted features are listed in Supplementary Tables S2, S4–S8, S10, S12, and S13. Missing values (NA) in UKB were mean-imputed for all features except oestradiol and rheumatoid factor, which were excluded from analysis due to a lack of data (oestradiol, 402.879 NAs; rheumatoid factor, 437.202 NAs). Cleaned features were min–max normalized (see “Data Processing”). Elevated liver enzymes (as a cutoff for “Patients at risk”) were defined as follows, with sex-specific cutoffs for males/females, respectively: AST (50/35 U/L), ALT (50/35 U/L), and γ-GT (60/40 U/L). Blood assays in AOU are not assessed at a predefined baseline time point but are passed for every hospital visit of the participants. Values were harmonized to standardize units and then filtered to remove measurements made in an emergency or urgent care setting. Outliers more than four standard deviations (SD) from the mean were removed to filter out likely erroneous values and then averaged over multiple instances in each year to reduce the number of missing data points.

Metabolomics

Utilizing NMR spectroscopy, the UKB characterized the plasma metabolomic/lipidomic signatures of 248,266 participants within the cohort (category 220). This high-throughput approach enabled the comprehensive quantification of a wide array of metabolites (n = 143), encompassing lipids, amino acids, and small-molecule intermediates, thereby providing an intricate biochemical snapshot pertinent to overall metabolic health. We extracted the 143 directly measured metabolomic features (see Supplementary Table S13; Fig. 2F); imputation and normalization procedures were performed in analogy to blood assays. All modeling analyses that included metabolomics data were performed only for the 248,266 participants for whom metabolomics data were available.

Genetic Data

Genetic profiling was conducted on a subset of 488,377 individuals within the UKB cohort, specifically targeting carriers of established hepatic risk variants. Eligible SNPs were collected via comprehensive manual literature research on variants associated with MASLD, liver disease, cirrhosis, or HCC (Fig. 2G; Supplementary Table S12). Genotypes were downloaded using the UKB gfetch utility. Specific genotype calls for hepatic risk variants were extracted with PLINK, including alignment and quality control. For each SNP, a linear regression model corrected for age, sex, and the first five principal components of the genotypes was performed. Our analysis stratified participants into noncarriers, heterozygous carriers, and homozygous carriers of the minor allele (refer to Supplementary Table S12 for details).

UKB Data Preprocessing

Data extraction and preprocessing steps were performed with RStudio (version 2023.12.1) and are publicly available in our GitHub repository (see “Data Availability”). Summary tables were created with the gtsummary package (93). Extracted data points are listed in Supplementary Tables. Data were processed separately for the following entities: initial assessment (demographics, lifestyle, self-reported diagnosis codes), EHRs, blood/serum, SNPs, and metabolomics. Data were treated as independent factor levels (categorical data) or imputed via mean imputation (UKB) or Multiple Imputation by Chained Equations (MICE) with classification and regression trees (CART; continuous data, AOU). Exact input features and missingness per column are displayed in Supplementary Table S10 (UKB) and Supplementary Table S28 (AOU). UKB patients were stratified into six folds according to the assessment center (ID 54-0.0), with folds 1 to 5 as the basis for the fivefold cross-validation, and fold 6 as the independent test set (Fig. 2B; Supplementary Tables S10 and S11). Outliers were limited to predefined physiologic limits to allow for shared scaling between UKB and AOU (Supplementary Table S12).

PheWAS

We matched incident HCC cases (C22.0) with controls via propensity score matching (1:10) on the covariates age, sex, and BMI. ICD Tenth Revision codes were converted to associated PheCodes (n = 1,799 PheCodes) using the PheWAS package in R. Beta and statistical significance were explored after Bonferroni correction for multiple testing. From ICD codes corresponding to significantly associated correlations (n = 137), a clinically meaningful and representative subset of ICD codes was then manually curated by two clinician authors (J. Clusmann, C.V. Schneider) to ensure comparability between the “All” and “PAR” cohorts (Supplementary Tables S5–S7, S9, and S22).

The AOU Cohort

The AOU cohort (RRID:SCR_027032) is a longitudinal cohort with 409,420 participants as of October 2024. The program collects and curates clinical, biological, and environmental determinants of health and disease, with procedures described in detail elsewhere (41). Health data were obtained through EHRs and participant surveys, which are available at www.researchallofus.org/data-tools/survey-explorer/. AOU gathers EHR data from >50 healthcare organizations; data stewards at provider organizations harmonize local data to the OMOP Common Data Model, which can then be accessed by researchers. For details on biospecimen collection and processing, visit https://allofus.nih.gov/funding-and-program-partners/biobank.

AOU Data Preprocessing

Preprocessing of tier 7 data in AOU was performed in line with UKB preprocessing, with the following exceptions: Alcohol consumption was approximated by frequency and amount of drinks consumed, with 14 g alcohol/drink as the average. While UKB invites for a primary assessment, AOU is linked to hospital stays, therefore providing fragmented data over longer time periods. Blood data were first aggregated (mean) per year to reduce NAs. For prospective evaluation, we then defined three anchor windows: <2010, 2010 to 2015, and 2015 to 2020. For each participant and window, we selected a single baseline year τ as the most recent calendar year with available blood data on or before the window start (e.g., for the 2010–2015 window, τ = 2010 if available; otherwise τ = 2009). The annual mean from year τ served as the laboratory input to the model for that window. Participants without any laboratory measurements on or before the window start did not contribute to that window. From τ, we then evaluated incident HCC within the subsequent 5 years. Participants with more than 25 NA values from the initial >100 data points were excluded from further analysis. Missing data were imputed with MICE using CART. AOU data were then converted to UKB units as documented in Supplementary Table S12.

Computational Infrastructure

Preprocessing and training of UKB data were performed on local workstations with Intel i7-i9 CPUs and 32 GB of RAM. Training times ranged between 10 minutes and 2 hours, without GPU requirements. The cross-validations were parallelized to use all but one CPU core available to the user.

Modeling and Hyperparameter Tuning

The modeling was conducted using scikit-learn version 1.2.2 (60). We chose the RFC and XGB as default model architectures based on literature and practicability. We split the UKB data into two sets: one for training (cross-validation) with exclusively England-based UKB assessment centers and one for testing the model performance (Glasgow, Edinburgh, Cardiff, Swansea, Newcastle). Training was carried out in a grouped fivefold cross-validation. Herein, each of the five folds (each consisting of 3 to 4 assessment centers; split, see Supplementary Tables S10 and S11) served for validation once and for training in four out of five models. A grid search evaluating different hyperparameter combinations in the fivefold cross-validation revealed an optimal architecture with 50 estimators and a max depth of 3. This architecture was then tested on the held-out UKB test set and the external validation set to avoid overfitting. All five models derived from cross-validation were integrated into a mean voting model as an ensemble classifier. This final model was then applied to the withheld test set, as well as the external test dataset from the AOU cohort.

Interpretability

Feature importance was estimated as the mean and SD of the accumulation of the impurity decrease within each tree via the scikit-learn library. SHapley Additive exPlanations (shap-library) were used for further interpretability with bootstrapping feature importance for individual participants, connecting optimal credit allocation with local explanations using the classic Shapley values from game theory and their related extensions (82, 94).

External Testing

External testing was performed in the research workbench of the AOU cohort (tier 7), with access only to the scripts, one hot encoder, and model object. Detailed scripts and requirements are available in our GitHub repository.

Statistical Analysis

Statistical comparisons between groups were performed using Welch’s two-sample t tests for continuous variables and χ² tests for categorical variables. For categorical variables with low cell counts (<20), the χ² approximation may be imprecise. Results are presented as mean ± SD for continuous variables and absolute numbers (percentages) for categorical variables, with values <20 masked for AOU to protect privacy as required by AOU. All statistical analyses were adjusted for multiple testing with false discovery rate or Bonferroni correction. For statistical comparison of AUROCs, two-sided DeLong tests were performed with Bonferroni post hoc correction. In-depth documentation for all statistical analyses can be found in our GitHub repository. Statistical tests were calculated in R (tables) or in Python (models).

Calibration

Calibration was conducted using cross-validation on the UKB training set. We employed Platt scaling (which fits a logistic regression model with a sigmoid function) as the primary calibration method, implemented through scikit-learn’s CalibratedClassifierCV with fivefold cross-validation stratified by assessment centers via the grouped structure of the training protocol. Calibration performance was evaluated using calibration slope, calibration intercept, and Brier score decomposition. Visual assessment was performed through calibration plots comparing observed versus predicted event rates across probability deciles. The calibrated models were subsequently deployed on the withheld UKB test set and the external All of Us validation cohort without refitting, allowing assessment of calibration transportability across different populations and healthcare systems.

Visualization

Visuals were created with RStudio 2024.04.0 as well as Python 3.9, including matplotlib, scikit-learn, seaborn (0.12.2), and shap (94), with all additional requirements being accessible via our GitHub repository. Figure assembly was performed with Inkscape 1.3.2, with integration of icons under a common license from Microsoft, Flaticons, Bioicons, and Healthicons.

Supplementary Material

Supplementary Tables ST1 - ST29

Supplementary Table 1 lists HCC case counts per UKB assessment center used in the study. Supplementary Table 2 lists the assessment-visit variables extracted and harmonized in UKB. Supplementary Table 3 reports PheWAS results used to derive disease-level features in UKB. Supplementary Table 4 lists the diagnosis-based inclusion criteria defining the PAR cohort in UKB. Supplementary Table 5 lists all single ICD code features used across UKB and AOU. Supplementary Table 6 lists grouped ICD (phecode/category) features used across UKB and AOU. Supplementary Table 7 maps self-reported conditions to ICD-based features in UKB. Supplementary Table 8 summarizes at-risk subset definitions and corresponding HCC/control sample sizes in UKB. Supplementary Table 9 provides the UKB ICD code summary representations. Supplementary Table 10 lists blood count and biochemistry features used for modeling in UKB and AOU. Supplementary Table 11 reports UKB blood count/biochemistry summary statistics stratified by HCC status and sex. Supplementary Table 12 lists genomic features (variants, coding, and availability) used in UKB modeling. Supplementary Table 13 lists metabolomics features used in UKB modeling. Supplementary Table 14 reports UKB metabolomics summary statistics stratified by HCC status. Supplementary Table 15 details the center-based train/test splits applied in UKB. Supplementary Table 16 provides UKB–AOU variable mappings and conversion summaries (min/max/median/mean and unit conversion where applicable). Supplementary Table 17 reports UKB threshold-independent performance metrics (AUROC, AUPRC) for all models and benchmarks. Supplementary Table 18 reports UKB threshold-dependent performance metrics across predefined operating points. Supplementary Table 19 reports UKB DeLong test results for AUROC comparisons between models/benchmarks. Supplementary Table 20 reports feature importances for Models A–E and TOP75–TOP15 in UKB. Supplementary Table 21 provides confusion matrices for selected thresholds in UKB and AOU (All and PAR where applicable). Supplementary Table 22 provides the AOU ICD code summary representations. Supplementary Table 23 reports AOU blood count/biochemistry summary statistics stratified by HCC status and sex. Supplementary Table 24 reports AOU threshold-independent performance metrics (AUROC, AUPRC) for all models and benchmarks. Supplementary Table 25 reports AOU threshold-dependent performance metrics across predefined operating points. Supplementary Table 26 reports AOU DeLong test results for AUROC comparisons between models/benchmarks. Supplementary Table 27 compares predicted risk distributions and related summary statistics across UKB and AOU. Supplementary Table 28 summarizes missingness (NA) patterns across features in AOU. Supplementary Table 29 provides an overview of all trained models, including modalities, feature sets, and evaluation cohorts (UKB/AOU).

Supplementary Figures S1 - S12

Supplementary Figure S1 shows participant inclusion and exclusion flowcharts for UKB and AOU, detailing data processing and cohort assembly steps. Supplementary Figure S2 shows age distributions at first HCC diagnosis in UKB and AOU and associations between metabolic biomarkers and HCC risk in PAR. Supplementary Figure S3 shows additional model performance metrics in UKB, including ROC curves, AUC comparisons across estimators, benchmark risk scores, and prediction score distributions. Supplementary Figure S4 shows the impact of missing-data thresholds on cohort size, disease prevalence, and model performance in UKB. Supplementary Figure S5 shows comparative performance of clinical models and AFP in the UKB proteomics subcohort using ROC and PRC analyses. Supplementary Figure S6 shows feature importance rankings for Models C and E across cohorts, including top-ranked features and aggregated feature group contributions. Supplementary Figure S7 shows calibration performance of the TOP15 model in the UKB All cohort before and after Platt scaling. Supplementary Figure S8 shows calibration performance of the TOP15 model in the UKB PAR cohort before and after Platt scaling. Supplementary Figure S9 shows additional external validation metrics in AOU, including prediction score distributions and confusion matrices across thresholds and sex strata. Supplementary Figure S10 shows the relationship between recall and NNS across thresholds in UKB and AOU for All and PAR cohorts. Supplementary Figure S11 shows calibration performance of the TOP15 model in the AOU All cohort before and after Platt scaling. Supplementary Figure S12 shows calibration performance of the TOP 15 model in the AOU PAR cohort before and after Platt scaling.

Data Availability

UKB data, including NMR metabolomics, are publicly available to bona fide researchers upon application at http://www.ukbiobank.ac.uk/using-the-resource/. Detailed information on predictors and endpoints used in this study is presented in Supplementary Tables S1–S29. This study used data from the AOU’s Controlled Tier Dataset version 7, available to authorized users on the Researcher Workbench. All code developed and used throughout this study has been made open source and is available on GitHub: https://github.com/schneiderlabac/PRE_SCREEN_HCC. A web calculator implementation of trained models is available at https://huggingface.co/spaces/schneiderlab/PRE_SCREEN_HCC.

Authors’ Disclosures

J. Clusmann reports personal fees from Johnson & Johnson outside the submitted work. O.S.M. El Nahhas reports other support from StratifAI GmbH outside the submitted work. A. Kaewdech reports grants and other support from Roche, Roche Diagnostics, and Abbott Laboratories and other support from Esai outside the submitted work. J. Calderaro reports other support from Clarapath and personal fees from Ipsen and Servier outside the submitted work. F. Tacke reports grants from AstraZeneca, MSD, Gilead, and Agomab and personal fees from Gilead, Falk, AstraZeneca, Boehringer, Madrigal, MSD, GSK, Ipsen, Mirum, Pfizer, Novo Nordisk, and Sanofi outside the submitted work. T. Bruns reports grants from the German Research Foundation during the conduct of the study, as well as personal fees from AdvanzPharma/Intercept Pharmaceuticals, SOBI, Novartis, Gilead, Falk Foundation, CSL Behring, Norgine, Intercept, AbbVie, Gilead, Merck, and Gore outside the submitted work. K.M. Schneider reports personal fees from AstraZeneca and Falk Foundation and other support from Ipsen outside the submitted work. J.N. Kather reports grants and personal fees from AstraZeneca; grants from GSK; other support from StratifAI GmbH, Synagen GmbH, and Spira Labs; personal fees from Owkin, DoMore Diagnostics, Bayer, Bioptimus, Bristol Myers Squibb, Daiichi Sankyo, Eisai Co. Ltd., F. Hoffmann-La Roche AG, Janssen Pharmaceuticals, Merck, Merck Sharp and Dohme, Mindpeak, MultiplexDx, Panakeia Technologies Ltd., and Pfizer outside the submitted work. C.V. Schneider reports grants and other support from Ipsen and personal fees from Takeda, AstraZeneca, Pfizer, and AirNA outside the submitted work. No disclosures were reported by the other authors.

Authors’ Contributions

J. Clusmann: Conceptualization, resources, data curation, software, formal analysis, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. P.-H. Koop: Data curation, software, validation, investigation, visualization, methodology, writing–review and editing. D.Y. Zhang: Data curation, validation, investigation, methodology. F. van Haag: Data curation, formal analysis, visualization, writing–review and editing, feedback. O.S.M. El Nahhas: Conceptualization, investigation, methodology, writing–review and editing, feedback. T. Seibel: Data curation, software, investigation, methodology, writing–review and editing, feedback. L. Žigutytė: Methodology, writing–review and editing, feedback. A. Kaewdech: Validation. J. Calderaro: Writing–review and editing, feedback. F. Tacke: Writing–review and editing, feedback. T. Luedde: Supervision, writing–review and editing, feedback. D. Truhn: Supervision, writing–review and editing, feedback. T. Bruns: Conceptualization, supervision, methodology, writing–review and editing, feedback. K.M. Schneider: Conceptualization, resources, software, supervision, funding acquisition, investigation, methodology, project administration, writing–review and editing, feedback. J.N. Kather: Conceptualization, resources, supervision, funding acquisition, validation, project administration, writing–review and editing, feedback. C.V. Schneider: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, methodology, writing–original draft, project administration, writing–review and editing.