Serologic Detection of Hepatocellular Carcinoma: Application of Machine Learning and Implications for Diagnostic...

Area of Science:

• Hepatology and Medical Oncology
• Biomarker Discovery and Validation
• Machine Learning in Healthcare

Background:

• The GALAD score, a biomarker-based model, aids in serologic diagnosis of hepatocellular carcinoma (HCC) but shows suboptimal performance in prospective studies.
• Real-world application of diagnostic models necessitates evaluation in prospective, real-world settings.
• Machine learning (ML) offers potential to enhance diagnostic accuracy for complex diseases like HCC.

Purpose of the Study:

• To apply machine learning (ML) to a large, prospectively collected HCC surveillance dataset.
• To develop and evaluate ML-based models for improved early detection of HCC.
• To compare the performance of ML models against the established GALAD score and individual biomarkers.

Main Methods:

• Trained two random forest (RF) models: GALAD-RF (using original GALAD variables) and RF-practical (using routine clinical/laboratory features).
• Evaluated models on a cohort of 3,473 patients with chronic liver disease, including 459 with HCC detected during surveillance (1998-2014).
• Compared model performance using area under the receiver operator curve (AUROC) and F1 score via repetitive cross-validation.

Main Results:

• The GALAD-RF model significantly outperformed the original GALAD logistic regression model in diagnostic accuracy (AUROC and F1 score).
• The RF-practical model also demonstrated superior performance compared to the original GALAD model.
• Both ML-based models showed better performance than individual biomarkers, and an online application was developed.

Conclusions:

• RF-based models enhance the diagnostic performance of the GALAD score within standard HCC surveillance programs.
• These ML models warrant further prospective validation for HCC early detection.
• Future studies could expand these models to predict HCC risk over specific timeframes.