AI-guided additive scoring model for differential diagnosis of primary liver cancer

Within the Liver Imaging Reporting and Data System (LI-RADS), LI-RADS Malignant (LR-M) lesions remain diagnostically challenging: imaging indicates malignancy, but often fails to distinguish hepatocellular carcinoma (HCC) from intrahepatic cholangiocarcinoma (iCCA). Established serum-based tools, such as GALAD, are optimized for HCC, but are not designed to resolve entity ambiguity within LR-M. In this pilot proof of concept study, we investigated whether small extracellular vesicles (EVs) integrated with routine serological markers could resolve LR-M lesion situation.

A rare LR-M cohort (HCC = 25, iCCA = 25) was evaluated using least absolute shrinkage and selection operator (LASSO) logistic regression and PCA to prioritize informative biomarkers. Hybrid models combined CD9⁺CD133/2⁺ and CD81⁺CD133/2⁺ EVs with alkaline phosphatase, serum C-reactive protein, CA19-9, and optional alpha-fetoprotein (AFP). Models were internally validated using an 80:20 train-test split and assessed with bootstrap and Monte Carlo perturbation (±5-20%).

Individual markers demonstrated limited discrimination (AUROC ≤0.82). Hybrid logistic regression models showed moderate internal discrimination (AUROC 0.86 with AFP; 0.91 without AFP). Translation into additive scoring systems using ROC-Youden-derived cut-offs yielded high internal AUROC estimates (e.g. 0.95-0.96), although these remain internally validated. A simplified 5-point PRISM score retained comparable discriminatory performance (AUROC ∼0.91). In an exploratory survival analysis among patients with iCCA (n = 25), those with above-median CD9⁺CD133/2⁺ EV levels (n = 13) had shorter overall survival (median 91 vs. 389 days; HR: 2.80, 95% CI: 1.16-6.74; p = 0.005).

Integrated EV and serological profiling could enable minimally invasive differentiation between HCC and iCCA within LR-M lesions. By transforming machine learning (ML)-guided model discovery into a clinically interpretable paper-and-pencil additive score, we illustrate a translational pathway from computational discovery to practical application.

This study provides a scientific rationale for integrating small EV phenotyping with conventional serum biomarkers to improve minimally invasive differentiation between HCC and iCCA in LR-M lesions. By translating a machine learning (ML)-derived hybrid model into a simple additive scoring system, we demonstrate how ML can yield clinically interpretable tools that bridge computational discovery and bedside application. The findings are particularly relevant for hepatologists, oncologists, and radiologists managing patients in whom imaging remains indeterminate and biopsy carries procedural risk. The 5-point model can be applied using routine laboratory values, providing a paper-and-pencil diagnostic aid that requires no specialized software or hardware. Although external multicenter validation is necessary given the cohort size, this approach illustrates how transparent, interpretable ML can support precision diagnostics in primary liver cancer.