从传统评分到人工智能：胰十二指肠切除术后胰瘘预测模型的演进与展望

From Traditional Scoring to Artificial Intelligence: Evolution and Prospects of Prediction Models for Postoperative Pancreatic Fistula after Pancreaticoduodenectomy

摘要: 胰十二指肠切除术(PD)是治疗胰头及壶腹周围肿瘤的主要手段，尽管围术期死亡率有所下降，但术后胰瘘(POPF)仍然是一个严重的并发症，临床相关性胰瘘(CR-POPF)可能导致高病死率以及医疗负担增加。本文详细回顾了PD术后胰瘘预测模型的发展历程，包括各个模型的具体改进和应用实例。从早期的传统评分系统，比如原始FRS、a-FRS和ua-FRS，到近年来应用的机器学习与深度学习模型，预测准确性逐渐提升。传统模型虽然简单实用，但在外部验证中常出现和报告结果不符的性能下降。随机森林、XGBoost、CatBoost、神经网络和深度学习等人工智能方法，结合了多方面临床变量、术后引流数据以及放射组学特征，内部验证的AUROC可超过0.80。部分联合模型在术前预测以及围术期管理中表现出优于传统FRS的潜力。然而，现在的研究仍以回顾性、单/少中心为主，外部验证数据不足，报告透明度以及泛化能力还较低。未来需开展大规模多中心前瞻性研究，推动模型融合、发展可解释AI、以及开发区域性模型。以实现从静态评分向动态、个体化精准风险分层工具的转变，最终优化围术期管理、降低CR-POPF相关严重结局。

Abstract: Pancreaticoduodenectomy (PD) is the primary therapeutic modality for periampullary and pancreatic head neoplasms. Although the perioperative mortality rate has declined, postoperative pancreatic fistula (POPF) remains a severe complication, and clinically relevant pancreatic fistula (CR-POPF) may lead to a high mortality rate and increased medical burden. This paper systematically reviews the developmental history of predictive models for POPF after PD. From early traditional scoring systems (e.g., the original Fistula Risk Score [FRS], adjusted FRS [a-FRS], and unadjusted FRS [ua-FRS]) to the machine learning and deep learning models widely applied in recent years, the predictive accuracy has been gradually improved. Despite their simplicity and practicability, traditional models often exhibit performance degradation in external validation. After integrating multi-dimensional clinical variables, postoperative dynamic drainage data and radiomic features, artificial intelligence methods (including random forest, XGBoost, CatBoost, neural networks and deep learning) have achieved an area under the receiver operating characteristic curve (AUROC) of over 0.80 in internal validation; some combined models outperform the traditional FRS and have demonstrated potential in preoperative prediction and perioperative management. However, current studies are still dominated by retrospective, single or small-sample multicenter research, with insufficient external validation data, low reporting transparency and poor generalization ability. In the future, large-scale multicenter prospective studies need to be conducted to promote model integration, develop explainable artificial intelligence, and establish regional predictive models. The ultimate goal is to realize the transformation from static scoring systems to dynamic and individualized precise risk stratification tools, thereby optimizing perioperative management and reducing CR-POPF-related severe outcomes.