机器学习在原发性肝癌术后复发预测中的应用与可解释性进展

Machine Learning in Predicting Postoperative Recurrence of Hepatocellular Carcinoma: Advances in Application and Explainability

摘要: 原发性肝癌(HCC)是世界上常见的恶性肿瘤之一，全球超过半数新发病例都发生在中国，手术切除是目前最主要的根治性治疗手段，但由于其术后复发率居高不下，严重影响患者的长期生存。因此，术后复发风险的精准预测对个体化治疗策略的制定具有至关重要的意义。近年来，随着人工智能的发展，机器学习(Machine learning)技术在肝癌的预后预测中展现出显著优势，其通过多模态数据融合与复杂模型构建显著提升了预测性能。然而，大多数先进机器学习模型具有“黑箱”特性，其决策过程不透明，限制了其在临床实践中的可信度与推广。本文系统综述了机器学习在肝癌术后复发预测中的应用进展，重点分析了基于临床数据、影像组学、病理图像等不同数据源的机器学习预测模型构建方法与性能，并深入探讨了以SHAP (Shapley Additive Explanations)为代表的可解释性技术在提升模型透明度及临床信任度方面的作用。文章进一步总结了当前研究面临的挑战，如数据异质性、模型泛化能力不足、可解释性标准缺失等，并展望了未来发展方向。

Abstract: Primary liver cancer (HCC) is one of the most common malignant tumors worldwide, with more than half of new cases occurring in China. Surgical resection is currently the primary curative treatment, but its high postoperative recurrence rate significantly affects long-term patient survival. Therefore, accurate prediction of postoperative recurrence risk is crucial for developing personalized treatment strategies. In recent years, with the advancement of artificial intelligence, machine learning (ML) technology has demonstrated significant advantages in prognostic prediction for liver cancer, notably enhancing predictive performance through multimodal data fusion and complex model construction. However, most advanced ML models exhibit “black box” characteristics, with opaque decision-making processes that limit their credibility and generalizability in clinical practice. This article systematically reviews the application progress of ML in postoperative recurrence prediction for liver cancer, focusing on the construction methods and performance of ML prediction models based on different data sources such as clinical data, radiomics, and pathological images. It also explores the role of interpretable techniques, represented by SHAP (Shapley Additive Explanations), in improving model transparency and clinical trust. The article further summarizes current research challenges, such as data heterogeneity, insufficient model generalization, and a lack of interpretability standards, and outlines future research directions.