从传统回归到机器学习：血液透析患者死亡风险预测模型的系统综述

From Traditional Regression to Machine Learning: A Systematic Review of Prediction Models for Mortality Risk in Hemodialysis Patients

摘要: 终末期肾病全球负担日益加重，血液透析患者预后异质性强，亟需精准预测模型以指导临床决策。本文系统梳理了血液透析预后预测模型的研究进展，涵盖传统回归模型、机器学习与动态预测、新型生物标志物及特化模型四大方向。研究发现，模型构建正从依赖传统指标向整合多维度数据与智能算法演进，显著提升了预测性能。然而，当前研究仍面临血管通路“风险悖论”、模型可解释性不足、外部验证性能衰减等争议与局限。未来应致力于融合因果推断与可解释AI技术，构建动态、可临床整合的下一代预测系统，推动预后管理向个体化、精准化发展。

Abstract: The global burden of end-stage renal disease is increasingly heavy, and the prognosis of hemodialysis patients is highly heterogeneous, creating an urgent need for accurate prediction models to guide clinical decision-making. This article systematically reviews the research progress in prognostic prediction models for hemodialysis, covering four main directions: traditional regression models, machine learning and dynamic prediction, novel biomarkers, and specialized models. The study finds that model development is evolving from reliance on traditional indicators towards the integration of multi-dimensional data and intelligent algorithms, which has significantly improved predictive performance. However, current research still faces controversies and limitations, such as the “risk paradox” of vascular access, insufficient model interpretability, and performance decay in external validation. Future efforts should be dedicated to integrating causal inference and explainable AI technologies to build the next generation of dynamic, clinically integrated prediction systems, thereby advancing prognosis management towards individuation and precision.