基于舌象的糖尿病相关预测模型系统评价

Tongue image-based prediction models for diabetes：A systematic review

关键词:  糖尿病  舌诊客观化  舌象  人工智能  机器学习  预测模型  系统评价

Key words:  diabetes mellitus  tongue diagnosis objective quantification  tongue image  artificial intelligence  machine learning  prediction model  systematic review

摘要:

[目的] 系统梳理基于舌象的糖尿病相关预测模型的研究进展，总结研究特点、技术路径及模型性能，并对现有文献进行质量评价，为未来研究提供参考。[方法] 系统检索中国知网（CNKI）、万方数据（Wanfang）、维普（VIP）、中国生物医学文献数据库（CBM）、PubMed、EMbase、Cochrane Library及Web of Science数据库平台相关文献，由2名研究者基于纳排标准独立筛选文献、提取资料，并采用PROBAST及PROBAST-AI工具评估偏倚风险与适用性。[结果] 最终纳入17篇文献，研究领域涵盖糖尿病诊断、并发症风险预测及中医证型辨识。模型多整合舌象与临床指标构建多模态模型，深度学习与集成学习是主流技术，多项模型展现出优秀的区分能力。然而，方法学质量整体中等，普遍缺乏独立的外部验证与校准度评估。绝大多数研究依赖单中心数据且缺乏标准化流程，模型泛化能力存疑。[结论] 基于舌象的糖尿病预测模型研究发展迅速，但临床转化受限，未来需通过严谨的研究设计、标准化信息采集、开展前瞻性外部验证等方式，推动该技术临床应用。

Abstract:

[Objective] To systematically review the research advances in diabetes-related predictive models based on tongue images，summarize their characteristics，technical approaches，and performance，evaluate the quality of existing literature，and provide a reference for future research. [Methods] We systematically searched relevant literature in Chinese National Knowledge Infrastructure（CNKI），Wanfang Data，China Biology Medicine disc（CBM），VIP Database，PubMed，Embase，Web of Science，and Cochrane Library. Two researchers independently screened the literature based on predefined inclusion/exclusion criteria and extracted data. The risk of bias and applicability were assessed using the PROBAST and PROBAST-AI tools. [Results] Seventeen articles were included. The research areas covered diabetes diagnosis，prediction of complication risks，and identification of Traditional Chinese Medicine syndrome patterns. Most models integrated tongue images with clinical indicators to construct multimodal models. Deep learning and ensemble learning were the mainstream techniques，with several models demonstrating excellent discriminatory ability. However，the overall methodological quality was moderate. All studies only conducted internal validation，severely lacking independent external validation and calibration assessment. The vast majority of studies relied on single-center data and lacked standardized procedures，casting doubt on the generalizability of the models. [Conclusion] Research on diabetes predictive models based on tongue images is developing rapidly. While current model performance appears promising，their clinical translation is limited. Future efforts should focus on rigorous study design，standardized information collection，and conducting prospective external validation to promote the clinical application of this technology.