维持性血液透析患者抑郁症状风险预测模型的系统评价——现状、局限与多模态建模的展望

Systematic Review of Risk Prediction Models for Depressive Symptoms in Maintenance Hemodialysis Patients—Current Status, Limitations and Prospects of Multimodal Modeling

摘要: 目的：系统评价维持性血液透析患者抑郁症状风险预测模型的研究现状、方法学质量、预测效能及局限性，并探讨多模态建模在该领域的应用前景，为后续构建更加精准、客观、可推广的预测工具提供依据。方法：系统的检索PubMed、Web of Science、Embase、Cochrane Library、CINHAL、知网、万方、维普和CBM的文献。检索时限为建库至2026年1月1日。两名研究者独立按CHARMS与PROBAST标准进行数据提取与偏倚风险评价，由于纳入研究数量较少且异质性较大，采用描述性分析总结模型构建、验证、性能及预测因子特征。结果：最终纳入6项研究，构建6个抑郁症状风险预测模型，均来自中国人群，其中5项为回顾性研究，1项为前瞻性队列研究。纳入研究中抑郁症状发生率为23.6%~45.37%。现有模型候选预测变量为20~34个，最终纳入预测因子多为4~5个，模型构建以Logistic回归为主，少数研究采用LASSO或SVM等方法；所有模型均仅进行了内部验证。模型区分度尚可，5项研究报告的AUC为0.680~0.882，2项研究报告的C-index分别为0.744和0.792。高频预测因子主要包括高龄、女性、长透析龄、低白蛋白、贫血、钙磷代谢紊乱及睡眠障碍。PROBAST评价显示，6项研究均为高偏倚风险，主要问题集中在样本代表性不足、缺失值处理报告不充分、结局评估依赖量表且缺乏盲法评定，以及缺乏独立外部验证。结论：现有MHD患者抑郁症状风险预测模型具有一定的临床筛查价值，但整体仍处于早期探索阶段，存在方法学质量不高、验证层级不足、泛化能力有限等问题。未来应在规范结局评估、扩大样本来源、加强外部验证的基础上，进一步探索人工智能支持下的多模态建模策略，以提升模型的准确性、客观性与临床可应用性。

Abstract: Objective: To systematically review the research status, methodological quality, predictive performance, and limitations of risk prediction models for depressive symptoms in maintenance hemodialysis (MHD) patients, and to explore the application prospects of multimodal modeling in this field, so as to provide a basis for the subsequent development of more accurate, objective, and generalizable prediction tools. Methods: A systematic literature search was conducted in PubMed, Web of Science, Embase, Cochrane Library, CINAHL, CNKI, Wanfang, VIP, and CBM from the inception of each database to January 1, 2026. Two researchers independently performed data extraction and risk of bias assessment in accordance with the CHARMS and PROBAST guidelines. Given the small number of included studies and high heterogeneity, a descriptive analysis was used to summarize the characteristics of model development, validation, performance, and predictors. Results: A total of 6 studies were finally included, involving 6 risk prediction models for depressive symptoms, all of which were derived from Chinese populations. Among them, 5 were retrospective studies and 1 was a prospective cohort study. The prevalence of depressive symptoms in the included studies ranged from 23.6% to 45.37%. The candidate predictive variables of the existing models numbered 20 to 34, with 4 to 5 predictors finally included in most models. Logistic regression was the dominant modeling method, while a few studies adopted LASSO or SVM. All models only underwent internal validation. The models showed acceptable discrimination: the AUC reported in 5 studies ranged from 0.680 to 0.882, and the C-index reported in 2 studies was 0.744 and 0.792, respectively. High-frequency predictors mainly included advanced age, female gender, long dialysis vintage, hypoalbuminemia, anemia, calcium-phosphorus metabolism disorders, and sleep disorders. According to the PROBAST assessment, all 6 studies were at high risk of bias, mainly due to inadequate sample representativeness, insufficient reporting of missing value handling, outcome assessment relying on scales without blinded evaluation, and lack of independent external validation. Conclusion: Current risk prediction models for depressive symptoms in MHD patients have certain clinical screening value, but they are still in the early exploratory stage overall, with problems such as low methodological quality, insufficient validation levels, and limited generalizability. In the future, on the basis of standardizing outcome assessment, expanding sample sources, and strengthening external validation, further exploration of artificial intelligence-assisted multimodal modeling strategies is warranted to improve the accuracy, objectivity, and clinical applicability of the models.