Interpretable Staging Prediction of Liver Cancer Based on Joint-Knowledge Network

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Radiology Reporting

Background:

• Accurate clinical staging of hepatocellular carcinoma (HCC) is vital for treatment and survival.
• Current staging methods face challenges in differentiating similar textual features and lack interpretability.

Purpose of the Study:

• To introduce KnowST, a joint-knowledge network for interpretable staging prediction in liver cancer.
• To leverage implicit knowledge in radiology reports for improved staging accuracy.

Main Methods:

• Developed a joint-knowledge network (KnowST) focusing on staging-related implicit knowledge.
• Utilized 'stages-to-stages' and 'factors-to-stages' relevance perspectives.
• Incorporated domain-specific word embeddings for Chinese radiology reports.
• Integrated tumor factor prediction for enhanced model credibility and interpretability.

Main Results:

• KnowST achieved 90.52% accuracy on 573 real-world radiology reports.
• Improved Area Under the Curve (AUC) by 7.69% compared to optimal baselines.
• Demonstrated superior stage identification and stable performance across multiple metrics.

Conclusions:

• KnowST effectively utilizes implicit information from radiology reports for interpretable clinical staging.
• The model enhances the credibility and interpretability of deep learning in staging prediction.
• KnowST offers a promising approach for improving hepatocellular carcinoma staging and patient outcomes.