基于机器学习模型的糖尿病患者低血糖预测在中国患者中的表现：一项系统评价与meta分析。
全文（开放获取） · 中文译文
摘要 目的：本研究旨在系统评估基于机器学习（ML）的模型对中国糖尿病患者低血糖的预测性能。 方法：我们系统检索了PubMed、Embase、Web of Science、Cochrane Library、CINAHL、CNKI和万方数据库，检索时间从建库至2026年2月。纳入研究聚焦于开发或验证基于机器学习（ML）的模型以预测中国糖尿病患者低血糖。研究筛选和数据提取由两名评审员独立完成。收集了研究特征、建模方法、预测因子、验证方法和模型性能的信息。使用随机效应模型汇总受试者工作特征曲线下面积（AUC）。采用预测模型偏倚风险评估工具（PROBAST）评估研究质量。 结果：共纳入13项研究，低血糖的合并发生率为25%（95% CI：17%-33%）。总体合并受试者工作特征曲线下面积（AUC）为0.90（95% CI：0.87-0.93）。按建模算法进行的亚组分析显示，极端梯度提升（XGBoost）的合并AUC为0.89，随机森林（RF）为0.88，支持向量机（SVM）为0.85，轻量梯度提升机（LightGBM）为0.84，逻辑回归（LR）为0.83，决策树（DT）模型为0.81。常见预测因子包括年龄、胰岛素使用、体质指数、HbA1c、肌酐和低血糖史。 结论：我们试图对基于机器学习的糖尿病患者低血糖预测模型进行全面概述。该领域的研究仍处于早期阶段，尽管已有一些具有良好区分性能的模型被报道。许多研究存在方法学局限和验证不足的问题。模型的稳健性和可解释性也令人担忧。需要付出更多努力来开发可靠且可解释的模型，并推动其在临床实践中的应用，以实现早期风险识别。 1 引言 糖尿病是全球最常见的慢性代谢性疾病之一[]。根据国际糖尿病联盟（IDF）糖尿病地图集，中国约有1.48亿成年人患有糖尿病，是全球糖尿病患者人数最多的国家[,]。中国拥有全球最大的糖尿病患者群体，其临床特征和糖尿病管理背景可能与西方国家报道的情况有所不同。中国2型糖尿病患者通常在相对较低的体质指数水平下发病，并表现出独特的代谢特征[]。此外，饮食模式、血糖监测可及性以及降糖药物使用[]的差异可能导致低血糖风险特征和预测变量分布的差异。 低血糖是降糖治疗期间最常见且可能未被充分认识的并发症之一[]。在接受胰岛素或口服降糖药的患者中，轻至中度低血糖的发生率据报告分别可达50%和45%，而重度低血糖分别发生于约21%和6%的患者[]。轻度低血糖可损害生活质量和治疗依从性，而重度低血糖与跌倒、心血管事件和死亡风险增加相关[,]。低血糖的报告病死率约为12.9%，在重度病例中升至24.9%[]。因此，早期识别高风险个体对于预防低血糖事件和改善糖尿病管理至关重要[]。 已开发出风险预测模型，以帮助临床医生识别易发生低血糖的患者。传统模型通常纳入人口统计学特征、临床变量和治疗相关变量[]。然而，它们的预测性能和跨人群的泛化能力仍然有限[]。由于糖尿病管理通常涉及复杂的治疗调整和异质性的患者特征，需要更精细的风险评估方法。随着复杂临床数据的可及性不断增加，基于机器学习（ML）的模型已被应用于低血糖预测。通过纳入多样化的临床信息，这些模型可以捕获复杂模式，并改善糖尿病护理中的风险分层[–]。先前的系统评价，例如Liu等人[]进行的系统评价，已总结了总体证据。然而，基于ML的预测模型在中国糖尿病患者中的性能和临床适用性尚未被评估。因此，本系统评价聚焦于中国糖尿病患者，旨在评估基于ML的低血糖预测模型的性能、方法学质量和临床适用性。 2 材料 2.1 方案与注册 本Meta分析已在PROSPERO注册（CRD420261338933），并按照诊断试验准确性系统评价和Meta分析优先报告条目（PRISMA-DTA）指南[]进行。 2.2 文献检索 在PubMed、Embase、Web of Science、Cochrane Library、CINAHL、中国知网（CNKI）和万方数据库中对英文和中文文献进行了全面检索。此外，对所有纳入研究的参考文献列表进行了手动筛选，以识别任何可能相关的文章。检索策略结合了与低血糖、糖尿病和基于机器学习（ML）的预测模型相关的医学主题词（MeSH）和自由文本关键词。主要检索词包括“hypoglycemia”、“diabetes mellitus”、“machine learning”、“artificial intelligence”、“deep learning”、“prediction model”、“risk factors”和“risk prediction”。使用布尔运算符（“OR”和“AND”）将不同概念组的检索词组合起来。检索覆盖了从数据库建库至2026年2月的研究，且仅考虑全文原始研究文章。所有检索到的记录均导入NoteExpress软件进行文献管理和去重。 2.3 纳入标准 我们采用Population、Index、Comparator、Outcome、Timing和Setting（PICOTS）框架[]来定义本综述的纳入标准：P（中国糖尿病患者的群体），I（预测模型），C（不适用），O（低血糖），T（糖尿病病程中的预测），以及S（临床环境，包括住院和门诊）。 纳入标准如下：[]涉及1型或2型糖尿病中国患者的研究；[]开发基于机器学习（ML）的低血糖预测模型的研究；[]报告受试者工作特征曲线下面积（AUC）作为模型性能衡量指标的研究；[]观察性研究设计（如队列研究或病例对照研究）；以及[]以英文或中文发表的文章。排除标准为：[]仅评估危险因素而未开发预测模型的研究；[]数据不完整或无法获取全文的研究；以及[]综述、社论、信件、病例报告、会议摘要或其他非原创研究。 2.4 研究筛选 两名评审员独立筛选所有检索到的记录的标题和摘要，以确定可能符合条件的研究并删除重复条目。然后根据预定义的纳入和排除标准评估剩余文章的全文是否符合条件。在全文筛选阶段记录排除原因。任何分歧均通过讨论解决，若无法达成共识，则咨询第三名评审员。 2.5 数据提取 两位作者（Yan和Song）使用基于CHARMS清单[]（预测建模研究系统评价的批判性评估和数据提取清单）制定的标准化数据提取表，独立提取了所有符合条件研究的数据。提取的信息包括第一作者、发表年份、研究设计、研究环境、数据集特征、建模方法、验证方法、模型性能和预测因子。所有提取的数据均与原始文章进行了交叉核对。任何差异均通过讨论解决，若无法达成共识，则咨询第三位评审者进行裁决。 采用预测模型偏倚风险评估工具（PROBAST）评估纳入研究的偏倚风险和适用性，该工具是一种为开发或验证预测模型的研究而设计的方法学工具 []。PROBAST评估四个领域的偏倚风险：参与者、预测变量、结局和分析。每个领域包含若干信号问题，这些问题被评定为“是/可能是”、“否/可能否”或“无信息”，这些回答用于判断该领域的偏倚风险为低、高或不明确。只有当所有信号问题均得到肯定回答时，该领域才被判定为低风险；否则，被认为具有高风险或不明确风险。适用性关注在三个领域——参与者、预测变量和结局——进行评估，依据与偏倚风险评估一致的标准 []。两名评价员独立进行质量评价，任何分歧通过讨论或与第三方评价员协商解决。 2.7 数据综合 本研究使用STATA 18.0软件进行 meta 分析。采用 Higgins I² 统计量评估纳入研究间的异质性。鉴于预测模型研究固有的异质性，进行了随机效应 meta 分析[]。以 AUC 作为预测性能的主要衡量指标，并报告合并估计值及相应的 95% 置信区间（CI）。AUC 在 0.7 至 0.9 之间被认为表明中等预测准确性，而 AUC 大于 0.9 则表明高预测准确性。当至少有2项研究可用时，根据相关研究特征进行亚组分析。通过逐一排除个别研究进行敏感性分析，以评估合并估计值的稳健性。采用 Egger 检验评估发表偏倚，并通过漏斗图进行视觉评估[]。双侧 P 值 < 0.05 被认为具有统计学显著性。 3 结果 3.1 研究选择 通过电子数据库检索共识别出5,450条记录，包括PubMed（n = 1,609）、Embase（n = 1,596）、Web of Science（n = 1,363）、CINAHL（n = 455）、CNKI（n = 415）和万方（n = 312）。去除3,215条重复记录后，剩余2,235篇文章进行标题和摘要筛选，其中2,143篇因不符合纳入标准而被排除。随后对92篇文章的全文进行合格性评估。其中，79项研究因以下原因被排除：未报告相关结局（n = 36）、研究人群不相关（n = 16）、以英语或中文以外的语言发表（n = 4）以及仅关注危险因素的分析（n = 23）。最终，13项研究纳入最终荟萃分析。研究筛选过程如PRISMA流程图所示（）。 文章筛选过程。 3.2 纳入原始研究的特征 本综述共纳入13项研究，探讨基于机器学习（ML）的糖尿病低血糖预测模型。这些研究发表于2022年至2026年间，且均在中国开展。大多数研究采用回顾性设计（n = 11），两项研究采用前瞻性设计。所纳入研究报告的低血糖患病率为4%至46.93%。大多数研究聚焦于2型糖尿病（T2DM）患者，而3项研究同时纳入了1型糖尿病（T1DM）和2型糖尿病人群。各研究的预测时限有所不同，范围从30分钟至12个月不等，其中大多数模型用于短期低血糖预测。大多数研究在单中心开展（n = 11），两项研究使用了多中心数据。所纳入研究的详细特征见。 所纳入文献的基本特征。 3.3 模型的开发与验证 所有纳入研究开发了基于机器学习（ML）的模型来预测低血糖。建模数据集的样本量范围为192至255,404名参与者，验证数据集范围为70至109,459名参与者。应用了多种建模方法，包括逻辑回归（LR）、随机森林（RF）、支持向量机（SVM）、决策树（DT）、极端梯度提升（XGBoost）、轻量梯度提升机（LightGBM）、深度神经网络（DNN）和长短期记忆（LSTM）网络。集成学习算法，特别是XGBoost和RF，经常被报道为最优模型，AUC值范围为0.822至0.978。七项研究进行了外部验证，而其余研究使用k折交叉验证或自举法进行了内部验证。关于区分度之外的模型性能评估，六项研究报告了使用校准曲线（n = 5）或Hosmer–Lemeshow检验（n = 1）进行校准评估。然而，所有纳入研究均未报告决策曲线分析或其他临床实用性 measures。关于模型特征、预测因子和验证的详细信息见。模型构建方法和预测性能。 注：DT = 决策树；DNN = 深度神经网络；ET = 极端随机树；GBM = 梯度提升机；GBDT = 梯度提升决策树；KNN = K-近邻；LightGBM = 轻量梯度提升机；LR = 逻辑回归；RF = 随机森林；SVM = 支持向量机；Xgboost = 极端梯度提升；* = 最优模型。 3.4 低血糖患病率 在纳入的13项研究中，12项报告了低血糖患病率并纳入荟萃分析。使用随机效应模型（REML），低血糖的合并患病率估计为25%（95% CI：17%–33%）。详细结果见森林图。 低血糖合并患病率的森林图。 3.5 纳入模型的预测因子 在纳入的13项研究中，所有研究均报告了其低血糖预测模型使用的预测因子。各项研究共提取了40多个不同的预测因子。为突出最相关的变量，图中仅呈现至少两项研究报告过的预测因子。最常纳入的预测因子为年龄、胰岛素使用、漏餐、BMI、HbA1c、肌酐和低血糖史，而其他预测因子在纳入模型中的分布汇总于图中。 基于机器学习低血糖模型中的预测因子。 3.6 荟萃分析结果 3.6.1 性能指标的荟萃分析。 森林图总结了纳入研究的AUC值。合并分析得出的总体AUC为0.90（95% CI: 0.87–0.93）（ and ）。根据建模算法进一步进行亚组分析。LightGBM的合并AUC为0.84（95% CI: 0.78–0.90; I2 = 86.75%, p < 0.001），XGBoost为0.89（95% CI: 0.83–0.95; I2 = 97.40%, p < 0.001），RF为0.88（95% CI: 0.84–0.93; I2 = 98.22%, p < 0.001），DT为0.81（95% CI: 0.75–0.88; I2 = 92.17%, p < 0.001），LR为0.83（95% CI: 0.81–0.86; I2 = 91.51%, p < 0.001），SVM为0.85（95% CI: 0.76–0.96; I2 = 98.53%, p < 0.001）（ and ）。 风险预测模型AUC的森林图。 RF和XGBoost合并AUC的森林图。 DT、LR、LightGBM和SVM模型的合并AUC森林图。 3.6.2 Meta分析亚组。 亚组分析按照研究设计、研究中心类型、预测变量类型和验证方法进行（）。按研究设计分层时，回顾性研究的合并曲线下面积（AUC）为0.88（95% CI：0.86–0.91；I2 = 96.07%），前瞻性研究的合并AUC为0.85（95% CI：0.80–0.89；I2 = 0.11%）。按研究中心类型分析时，单中心研究的合并AUC为0.88（95% CI：0.85–0.91；I2 = 86.03%），多中心研究的合并AUC为0.87（95% CI：0.87–0.88；I2 = 59.42%）。按预测变量类型分层时，使用临床变量开发的模型显示合并AUC为0.88（95% CI：0.85–0.90；I2 = 88.93%），而基于持续葡萄糖监测（CGM）时间序列数据的模型合并AUC为0.89（95% CI：0.85–0.94；I2 = 95.50%）。根据验证方法，采用内部验证的模型合并AUC为0.90（95% CI：0.86–0.93；I2 = 78.53%），而采用外部验证的模型合并AUC为0.87（95% CI：0.85–0.89；I2 = 95.70%）（）。 低血糖预测模型合并AUC的亚组分析。 3.7 纳入研究的质量评估 使用PROBAST工具评估纳入研究的偏倚风险及适用性()。在纳入的13项研究中，4项研究总体偏倚风险较低，2项研究偏倚风险不明确，其余研究偏倚风险较高。在领域层面评估中，大多数研究在受试者领域和预测变量领域偏倚风险较低，但部分研究在结局领域和分析领域存在问题。在适用性方面，大多数研究在受试者领域和预测变量领域显示出较低的担忧，少数研究在结局领域存在高或不明确的担忧。 偏倚风险评估。 ROB = 偏倚风险。 “+”表示偏倚风险低或适用性担忧低； “−”表示偏倚风险高或适用性担忧高； “?”表示偏倚风险不明确或适用性担忧不明确。 3.8 敏感性分析和发表偏倚 敏感性分析采用逐项排除法进行，当逐一排除每项研究后，合并结果未发生显著变化，这表明没有哪一项研究对总体估计值产生了显著影响 ()。采用漏斗图目测检查和Begg检验来评估发表偏倚。Begg检验给出的P值为0.669，表明没有统计学上显著的发表偏倚证据。大多数研究位于预期的漏斗区域，漏斗图几乎对称，表明本Meta分析中无明显发表偏倚 ()。 敏感性分析。 漏斗图检查。 4 讨论 4.1 预测性能与方法学 随着人口快速老龄化和饮食结构的改变，近年来中国糖尿病的患病率大幅上升[]，导致低血糖及其相关不良结局的负担日益加重[]。在本荟萃分析中，中国糖尿病患者中低血糖的合并患病率为25%（95% CI：17%–33%），凸显了该人群中低血糖的临床负担。中国拥有全球最大的糖尿病患者人群，且大多数患者为2型糖尿病（T2DM）。鉴于中国糖尿病患者群体规模庞大且具有特定的临床和医疗保健背景，在该人群中开发或验证的预测模型可能为低血糖风险评估和临床应用提供更相关的证据。与此同时，人工智能的进步促进了大量旨在识别低血糖高危患者的预测模型的开发[]。据我们所知，这项研究代表
全文（开放获取） · 英文原文
【Abstract】 OBJECTIVES: This study aimed to systematically evaluate the predictive performance of machine learning (ML)-based models for predicting hypoglycemia in Chinese patients with diabetes. METHODS: We systematically searched PubMed, Embase, Web of Science, the Cochrane Library, CINAHL, CNKI, and Wanfang databases from inception to February 2026. Eligible studies focused on the development or validation of ML-based models for predicting hypoglycemia in Chinese patients with diabetes. Study selection and data extraction were performed independently by two reviewers. Information on study characteristics, modeling approaches, predictors, validation methods, and model performance was collected. The area under the receiver operating characteristic curve (AUC) was synthesized using a random-effects model. Study quality was assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST). RESULTS: A total of 13 studies were included, and the pooled prevalence of hypoglycemia was 25% (95% CI: 17%-33%). The overall pooled area under the receiver operating characteristic curve (AUC) was 0.90 (95% CI: 0.87-0.93). Subgroup analyses by modeling algorithms showed pooled AUCs of 0.89 for extreme gradient boosting (XGBoost), 0.88 for random forest (RF), 0.85 for support vector machine (SVM), 0.84 for Light Gradient Boosting Machine (LightGBM), 0.83 for logistic regression (LR), and 0.81 for decision tree (DT) models. Common predictors included age, insulin use, body mass index, HbA1c, creatinine, and history of hypoglycemia. CONCLUSION: We attempted to provide a comprehensive overview of machine learning-based prediction models for hypoglycemia in patients with diabetes. Research in this field remains at an early stage, although several models with good discriminatory performance have been reported. Methodological limitations and insufficient validation were observed in many studies. Concerns regarding model robustness and interpretability also exist. More efforts to develop reliable and interpretable models and to promote their application in clinical practice for early risk identification are needed. 【1 Introduction】 Diabetes mellitus is one of the most prevalent chronic metabolic diseases worldwide []. According to the International Diabetes Federation (IDF) Diabetes Atlas, approximately 148 million adults in China are living with diabetes, representing the largest population of individuals with diabetes worldwide [,]. China has the largest population of individuals with diabetes worldwide, with clinical characteristics and diabetes management contexts that may differ from those reported in Western countries. Chinese patients with type 2 diabetes often develop diabetes at relatively lower BMI levels and exhibit distinct metabolic characteristics []. In addition, differences in dietary patterns, glucose monitoring accessibility, and glucose-lowering medication use [] may contribute to variations in hypoglycemia risk profiles and predictive variable distributions. Hypoglycemia is one of the most common and potentially underrecognized complications during glucose-lowering therapy []. Among patients receiving insulin or oral hypoglycemic agents, the incidence of mild-to-moderate hypoglycemia has been reported to reach 50% and 45%, respectively, while severe hypoglycemia occurs in approximately 21% and 6% of these patients []. Mild hypoglycemia can impair quality of life and treatment adherence, whereas severe hypoglycemia is associated with increased risks of falls, cardiovascular events, and mortality [,]. The reported case fatality rate of hypoglycemia is approximately 12.9%, rising to 24.9% in severe cases []. Early identification of individuals at high risk is therefore essential to prevent hypoglycemic events and improve diabetes management []. Risk prediction models have been developed to assist clinicians in identifying patients vulnerable to hypoglycemia. Conventional models generally incorporate demographic characteristics, clinical, and treatment-related variables []. However, their predictive performance and generalizability across populations remain limited []. As diabetes management often involves complex treatment adjustments and heterogeneous patient characteristics, more refined approaches to risk assessment are needed. With the increasing availability of complex clinical data, machine learning (ML)–based models have been applied to hypoglycemia prediction. By incorporating diverse clinical information, these models can capture complex patterns and improve risk stratification in diabetes care [–]. Previous systematic reviews, such as that conducted by Liu et al.[], have summarized the overall evidence. However, the performance and clinical applicability of ML-based prediction models among Chinese patients with diabetes have not been evaluated. Therefore, this systematic review focuses on Chinese patients with diabetes to to evaluate the performance, methodological quality, and clinical applicability of ML-based prediction models for hypoglycemia. 【2 Material】 【2.1 Protocol and registration】 This meta-analysis was registered in PROSPERO (CRD420261338933) and conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy Studies (PRISMA-DTA) guidelines []. 【2.2 Literature search】 A comprehensive search of both English and Chinese literature was conducted in the databases PubMed, Embase, Web of Science, the Cochrane Library, CINAHL, China National Knowledge Infrastructure (CNKI), and the Wanfang Database. In addition, the reference lists of all included studies were manually screened to identify any potentially relevant articles. The search strategy combined Medical Subject Headings (MeSH) terms and free-text keywords related to hypoglycemia, diabetes, and ML –based prediction models. Key search terms included “hypoglycemia”, “diabetes mellitus”, “machine learning”, “artificial intelligence”, “deep learning”, “prediction model”, “risk factors”, and “risk prediction”. Boolean operators (“OR” and “AND”) were used to combine search terms across different concept groups. The search covered studies from database inception to February 2026, and only full-text original research articles were considered. All retrieved records were imported into NoteExpress software for literature management and duplicate removal. 【2.3 Eligibility criteria】 We applied the Population, Index, Comparator, Outcome, Timing, and Setting (PICOTS) framework [] to define the eligibility criteria for this review: P (Chinese patients with diabetes mellitus), I (Prediction models), C (not applicable), O (hypoglycemia), T (prediction during the course of diabetes), and S (clinical settings, including inpatient and outpatient care). The inclusion criteria were as follows: [] studies involving Chinese patients with type 1 or type 2 diabetes; [] studies developing ML–based models for hypoglycemia prediction; [] studies reporting the area under the receiver operating characteristic curve (AUC) as a measure of model performance; [] observational study designs (e.g., cohort or case-control studies); and [] articles published in English or Chinese. The exclusion criteria were: [] studies assessing risk factors without developing prediction models; [] studies with incomplete data or unavailable full texts; and [] reviews, editorials, letters, case reports, conference abstracts, or other non-original studies. 【2.4 Study selection】 Two reviewers independently screened the titles and abstracts of all retrieved records to identify potentially eligible studies and remove duplicate entries. The full texts of the remaining articles were then assessed for eligibility according to the predefined inclusion and exclusion criteria. Reasons for exclusion were documented during the full-text screening stage. Any disagreements were resolved through discussion, and a third reviewer was consulted when consensus could not be reached. 【2.5 Data extraction】 Two authors (Yan and Song) independently extracted data from all eligible studies using a standardized data extraction form developed based on the CHARMS checklist [](Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies). The extracted information included the first author, publication year, study design, study setting, dataset characteristics, modeling methods, validation approach, model performance, and predictors. All extracted data were cross-checked against the original articles. Any discrepancies were resolved through discussion, and if consensus could not be reached, a third reviewer was consulted for adjudication. 【2.6 Quality appraisal】 The risk of bias and applicability of the included studies were evaluated using the Prediction Model Risk of Bias Assessment Tool (PROBAST), a methodological instrument designed for studies developing or validating prediction models []. PROBAST assesses risk of bias across four domains: participants, predictors, outcome, and analysis. Each domain contains several signaling questions that are rated as “yes/probably yes,” “no/probably no,” or “no information,” and the responses are used to determine whether the domain has a low, high, or unclear risk of bias. A domain was judged to be at low risk only when all signaling questions were answered positively; otherwise, it was considered to have high or unclear risk. Applicability concerns were evaluated in three domains—participants, predictors, and outcome—based on criteria aligned with the risk-of-bias assessment []. Two reviewers independently conducted the quality assessment, and any disagreements were resolved through discussion or consultation with a third reviewer. 【2.7 Data synthesis】 In this study, STATA 18.0 software was used to conduct the meta-analysis. Heterogeneity among the included studies was assessed using the Higgins I² statistic. Given the inherent heterogeneity of prediction modeling studies, a random-effects meta-analysis was performed []. The AUC was used as the primary measure of predictive performance, and pooled estimates were reported with corresponding 95% confidence intervals(CI). An AUC between 0.7 and 0.9 was considered to indicate moderate predictive accuracy, while an AUC greater than 0.9 indicated high predictive accuracy. Subgroup analyses were performed based on relevant study characteristics when at least two studies were available. Sensitivity analysis was conducted by sequentially excluding individual studies to assess the robustness of the pooled estimates. Publication bias was evaluated using Egger’s test and visually assessed with funnel plots []. A two-sided P value < 0.05 was considered statistically significant. 【3 Results】 【3.1 Study selection】 A total of 5,450 records were identified through electronic database searches, including PubMed (n = 1,609), Embase (n = 1,596), Web of Science (n = 1,363), CINAHL (n = 455), CNKI (n = 415), and Wanfang (n = 312). After removing 3,215 duplicate records, 2,235 articles remained for title and abstract screening, of which 2,143 were excluded for not meeting the inclusion criteria. The full texts of 92 articles were subsequently assessed for eligibility. Among them, 79 studies were excluded for the following reasons: not reporting relevant outcomes (n = 36), irrelevant study populations (n = 16), publication in languages other than English or Chinese (n = 4), and analyses focusing only on risk factors (n = 23). Ultimately, 13 studies were included in the final meta-analysis. The study selection process is illustrated in the PRISMA flow diagram (). 【The article selection process.】 【3.2 Characteristics of the included primary studies】 A total of 13 studies investigating ML–based prediction models for hypoglycemia in patients with diabetes were included in this review. The studies were published between 2022 and 2026 and were all conducted in China. Most studies used a retrospective design (n = 11), while two studies adopted a prospective design. The prevalence of hypoglycemia reported in the included studies ranged from 4% to 46.93%. The majority of studies focused on patients with type 2 diabetes mellitus (T2DM), whereas 3 studies included both type 1 diabetes mellitus (T1DM) and T2DM populations. Prediction horizons varied across studies, ranging from 30 minutes to 12 months, with most models designed for short-term hypoglycemia prediction. Most studies were conducted in single-center settings (n = 11), while two studies used multicenter data. Detailed characteristics of the included studies are presented in . 【Basic characteristics of the included literature.】 【3.3 Model development and validation】 All included studies developed ML-based models to predict hypoglycemia. The sample sizes of the modeling datasets ranged from 192 to 255,404 participants, with validation datasets ranging from 70 to 109,459 participants. A variety of modeling approaches were applied, including logistic regression (LR), random forest (RF), support vector machine (SVM), decision tree (DT), extreme gradient boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), deep neural networks (DNN), and long short-term memory (LSTM) networks. Ensemble learning algorithms, particularly XGBoost and RF, were frequently reported as the optimal models, with AUC values ranging from 0.822 to 0.978. Seven studies conducted external validation, whereas the remaining studies performed internal validation using k-fold cross-validation or bootstrapping. Regarding model performance evaluation beyond discrimination, six studies reported calibration assessment using calibration curves (n = 5) or the Hosmer–Lemeshow test (n = 1). However, none of the included studies reported decision-curve analysis or other measures of clinical utility. Detailed information on model characteristics, predictors, and validation is provided in . 【Model construction method and predictive performance.】 Note: DT = Decision Tree; DNN = Deep neural network; ET = Extra Trees; GBM = gradient boosting machine; GBDT = gradient boosting decision tree; KNN = K-Nearest Neighbors; LightGBM = Light Gradient Boosting Machine; LR = Logistic Regression; RF = Random Forest; SVM = Support Vector Machine; Xgboost = Extreme Gradient Boosting; * = optimal model. 【3.4 Prevalence of hypoglycemia】 Among the 13 included studies, 12 reported the prevalence of hypoglycemia and were included in the meta-analysis. Using a random-effects model (REML), the pooled prevalence of hypoglycemia was estimated to be 25% (95% CI: 17%–33%). The detailed results are presented in the forest plot (). 【Forest plots for the pooled prevalence of hypoglycemia.】 【3.5 Predictors of the included models】 Among the 13 included studies, all reported the predictors used in their hypoglycemia prediction models. Across the studies, more than 40 unique predictors were extracted. To highlight the most relevant variables, presents only those predictors that were reported in at least two studies. The most frequently included predictors were age, insulin use, Meal omission, BMI, HbA1c, creatinine, and history of hypoglycemia, while the distribution of other predictors across the included models is summarized in the . 【Predictors in ML-based hypoglycemia models.】 【3.6 Results of meta-analysis】 【3.6.1 Meta-analysis of performance measures.】 The forest plot summarizes the AUC values of the included studies. The pooled analysis yielded an overall AUC of 0.90 (95% CI: 0.87–0.93) ( and ). Subgroup analyses were further conducted according to modeling algorithms. The pooled AUCs were 0.84 (95% CI: 0.78–0.90; I2 = 86.75%, p < 0.001) for LightGBM, 0.89 (95% CI: 0.83–0.95; I2 = 97.40%, p < 0.001) for XGBoost, 0.88 (95% CI: 0.84–0.93; I2 = 98.22%, p < 0.001) for RF, 0.81 (95% CI: 0.75–0.88; I2 = 92.17%, p < 0.001) for DT, 0.83 (95% CI: 0.81–0.86; I2 = 91.51%, p < 0.001) for LR, and 0.85 (95% CI: 0.76–0.96; I2 = 98.53%, p < 0.001) for SVM ( and ). 【Forest plot of the AUC for the risk prediction model.】 【Forest Plot of Pooled AUCs for RF and XGBoost.】 【Forest Plot of Pooled AUCs for DT, LR, LightGBM, and SVM Models.】 【3.6.2 Subgroup of meta-analysis.】 The subgroup analyses were done in accordance to study design, center type, predictor type and validation method (). When stratified by study design, the pooled AUC was 0.88 (95% CI: 0.86–0.91; I2 = 96.07%) for retrospective studies and 0.85 (95% CI: 0.80–0.89; I2 = 0.11%) for prospective studies. In the analysis by center type, the pooled AUC was 0.88 (95% CI: 0.85–0.91; I2 = 86.03%) for single-center studies and 0.87 (95% CI: 0.87–0.88; I2 = 59.42%) for multicenter studies. When stratified by predictor type, models developed using clinical variables showed a pooled AUC of 0.88 (95% CI: 0.85–0.90; I2 = 88.93%), whereas models based on CGM time-series data yielded a pooled AUC of 0.89 (95% CI: 0.85–0.94; I2 = 95.50%). According to validation method, models with internal validation demonstrated a pooled AUC of 0.90 (95% CI: 0.86–0.93; I2 = 78.53%), while those with external validation showed a pooled AUC of 0.87 (95% CI: 0.85–0.89; I2 = 95.70%) (). 【Subgroup Analyses of Pooled AUCs for Hypoglycemia Prediction Models.】 【3.7 Quality assessment of included studies】 A PROBAST tool was used to evaluate the risk of bias and applicability of the included studies (). Out of 13 included studies, 4 studies were found to be at low overall risk of bias, 2 were found to be at unclear risk of bias, and the rest of the studies were found to be at high risk of bias. In the domain level assessment, most studies had low risk of bias in participants and predictors domains, but some studies had issues in outcome and analyses domains. Concerning applicability, most studies presented low concerns in participants and predictors domains, and a small number of studies had high or unclear concerns in outcome domain. 【Risk of bias assessment.】 ROB = risk of bias. “+” indicates low risk of bias or low concern regarding applicability; “−” indicates high risk of bias or high concern regarding applicability; “?” indicates unclear risk of bias or unclear concern regarding applicability. 【3.8 Sensitivity analysis and publication bias】 The sensitivity analysis was performed by the leave-one-out approach and the pooled findings did not significantly change when each study was sequentially excluded which means that there is no one study that has played a significant role in the overall estimate (). Visual inspection of the funnel plot and Begg test were used to evaluate publication bias. Begg test gave a P value of 0.669, which indicates no statistically significant evidence of publication bias. Most of the studies were found to be located at the expected funnel area, and the funnel plot was almost symmetrical, indicating no obvious presence of publication bias in this meta-analysis (). 【Sensitivity Analysis.】 【Funnel plot examination.】 【4 Discussion】 【4.1 Predictive performance and methodology】 With the rapid aging of the population and changes in dietary patterns, the prevalence of diabetes in China has increased substantially in recent years [], leading to a growing burden of hypoglycemia and its associated adverse outcomes []. In the present meta-analysis, the pooled prevalence of hypoglycemia among Chinese patients with diabetes was 25% (95% CI: 17%–33%), highlighting the clinical burden of hypoglycemia in this population. China has the largest population of individuals with diabetes worldwide, and most affected patients have type 2 diabetes mellitus (T2DM). Given the large population size and specific clinical and healthcare contexts of Chinese patients with diabetes, prediction models developed or validated in this population may provide more relevant evidence for hypoglycemia risk assessment and clinical application. In parallel, advances in artificial intelligence have facilitated the development of numerous prediction models aimed at identifying patients at high risk of hypoglycemia []. To our knowledge, this study represent