Explainable Machine Learning Accurately Classifies Liver Cancer Using Routine Clinical Data

A transparent machine learning framework based on routine clinical and biochemical data can accurately classify hepatocellular carcinoma while providing interpretable explanations of model behavior, according to a diagnostic evaluation conducted in a Vietnamese patient cohort. Hepatocellular carcinoma represents one of the leading causes of cancer-related mortality globally and stands among the primary causes of cancer-related deaths in Vietnam.

The study utilized the VD_HCC dataset, an institutional cohort of 880 Vietnamese patients collected at a Vietnamese healthcare center. The clinical sample comprised 303 individuals diagnosed with biopsy-confirmed hepatocellular carcinoma alongside 577 non-hepatocellular carcinoma patients. To prepare these routine clinical and laboratory records for algorithmic analysis, the investigators applied a specialized pre-processing pipeline structured to ensure data consistency, minimize background noise, and improve downstream classification performance.

To isolate the most informative diagnostic biomarkers from the broader dataset, the researchers evaluated eight distinct feature selection methods. Biochemical and clinical features that were consistently selected across multiple computational techniques were subsequently integrated with clinical domain expertise to construct optimized feature subsets. Multiple classification algorithms, emphasizing ensemble learning methods such as XGBoost, AdaBoost, and LightGBM, were systematically trained and benchmarked across these engineered biomarker subsets. Among the evaluated configurations, an optimized ten-variable collection designated as Set 5 demonstrated superior diagnostic capability, with alpha-fetoprotein, hepatitis B surface antigen, carcinoembryonic antigen, aspartate aminotransferase, platelet count, and alanine aminotransferase emerging as the leading predictive contributors.

Across 30 repeated experimental runs executed with independent training and test data partitions, the XGBoost model utilizing the ten-feature Set 5 achieved a mean diagnostic accuracy of 0.938 with a standard deviation of 0.018, alongside a mean F1-score of 0.909 with a standard deviation of 0.022. When evaluated on a locked hold-out test set, both the XGBoost and LightGBM ensemble architectures achieved identical peak receiver operating characteristic area under the curve values of 0.979. In precision-recall space on the locked hold-out partition, the XGBoost classifier achieved an area under the precision-recall curve of 0.971, while the LightGBM classifier recorded a corresponding precision-recall area under the curve value of 0.965.

To determine model resilience against severe epidemiological skew, the researchers subjected the ensemble models to an imbalanced stress-test scenario reflecting an extreme ratio of one hepatocellular carcinoma case to twenty-one non-carcinoma cases. Under this heavy 1:21 imbalance stress test, XGBoost and LightGBM maintained comparable overall classification accuracy. XGBoost demonstrated higher precision, recall, F1-score, and receiver operating characteristic area under the curve metrics under the skewed condition. Meanwhile, LightGBM recorded a slightly higher precision-recall area under the curve of 0.878 with a standard deviation of 0.089 compared with 0.861 and a standard deviation of 0.093 for XGBoost, both substantially outperforming the baseline no-skill precision-recall threshold of 0.045.

Model interpretability was evaluated using a two-layer explainable artificial intelligence analysis combining SHapley Additive exPlanations, Local Interpretable Model-agnostic Explanations, Partial Dependence Plots, and an empirical feature-correlation graph. This dual-layer interpretability strategy allowed the investigators to compare individual feature attributions directly and inspect the fitted behavior of the predictive algorithms. For the final Set 5 XGBoost architecture, aggregated Local Interpretable Model-agnostic Explanations and SHapley Additive exPlanations produced exceptionally concordant biomarker importance hierarchies, demonstrating a Spearman rank correlation coefficient rho of 0.976 and a top-six Jaccard similarity index of 1.000, while the feature-correlation graph offered a complementary characterization of interactions across the clinical variables.

The study authors emphasized critical boundaries regarding the interpretation of their computational findings, noting that the explainable artificial intelligence outputs describe the internal mechanics of the fitted model and statistical associations present in the study data, rather than establishing causal medical mechanisms or independently validated clinical importance. Overall, the investigators concluded that ensemble classifiers retain relatively stable discriminative performance across escalating imbalance conditions despite severe class disparity. The proposed framework provides an interpretable perspective on hepatocellular carcinoma classification by systematically distinguishing the specific routine biochemical markers that contribute most toward separating malignant cases from non-malignant controls.

Subject of Research: Diagnostic classification of hepatocellular carcinoma using explainable machine learning

Article Title: Diagnostic classification of hepatocellular carcinoma in the Vietnamese population using clinical and biochemical data: explainable machine learning approaches

Article References: Pham, T. T., Pham, T. H. T., Nguyen, H. P., Pham, Q. T., Le, V. K. A., Pham, G. A., Tran, B. G., Luu, H. N., Vu, D. H., Tran, T. H., Le, T. L., & Vu, H. (2026). Diagnostic classification of hepatocellular carcinoma in the Vietnamese population using clinical and biochemical data: explainable machine learning approaches. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03916-0

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03916-0

Keywords: hepatocellular carcinoma, machine learning, explainable artificial intelligence, biomarkers, XGBoost, LightGBM, cancer diagnosis, Vietnam, Diagnostic, classification, hepatocellular, carcinoma