University College London, London, UK.
British Heart Foundation Data Science Centre, Health Data Research UK, London, UK.
BMJ Open Diabetes Res Care. 2024 Jun 4;12(3):e004191. doi: 10.1136/bmjdrc-2024-004191.
None of the studies of type 2 diabetes (T2D) subtyping to date have used linked population-level data for incident and prevalent T2D, incorporating a diverse set of variables, explainable methods for cluster characterization, or adhered to an established framework. We aimed to develop and validate machine learning (ML)-informed subtypes for type 2 diabetes mellitus (T2D) using nationally representative data.
In population-based electronic health records (2006-2020; Clinical Practice Research Datalink) in individuals ≥18 years with incident T2D (n=420 448), we included factors (n=3787), including demography, history, examination, biomarkers and medications. Using a published framework, we identified subtypes through nine unsupervised ML methods (K-means, K-means++, K-mode, K-prototype, mini-batch, agglomerative hierarchical clustering, Birch, Gaussian mixture models, and consensus clustering). We characterized clusters using intracluster distributions and explainable artificial intelligence (AI) techniques. We evaluated subtypes for (1) internal validity (within dataset; across methods); (2) prognostic validity (prediction for 5-year all-cause mortality, hospitalization and new chronic diseases); and (3) medication burden.
: We identified four T2D subtypes: metabolic, early onset, late onset and cardiometabolic. : Subtypes were predicted with high accuracy (F1 score >0.98). : 5-year all-cause mortality, hospitalization, new chronic disease incidence and medication burden differed across T2D subtypes. Compared with the metabolic subtype, 5-year risks of mortality and hospitalization in incident T2D were highest in late-onset subtype (HR 1.95, 1.85-2.05 and 1.66, 1.58-1.75) and lowest in early-onset subtype (1.18, 1.11-1.27 and 0.85, 0.80-0.90). Incidence of chronic diseases was highest in late-onset subtype and lowest in early-onset subtype. : Compared with the metabolic subtype, after adjusting for age, sex, and pre-T2D medications, late-onset subtype (1.31, 1.28-1.35) and early-onset subtype (0.83, 0.81-0.85) were most and least likely, respectively, to be prescribed medications within 5 years following T2D onset.
In the largest study using ML to date in incident T2D, we identified four distinct subtypes, with potential future implications for etiology, therapeutics, and risk prediction.
迄今为止，尚无研究对 2 型糖尿病（T2D）进行亚型分类，这些研究都没有使用与发病和现患 T2D 相关的人群水平数据，也没有纳入多样化的变量、可解释的聚类特征描述方法，或遵循既定框架。我们旨在使用全国代表性数据，开发和验证基于机器学习（ML）的 2 型糖尿病（T2D）亚型。
我们对基于人群的电子健康记录（2006-2020 年；临床实践研究数据链接）中≥18 岁的新发 T2D 患者（n=420448）进行分析，这些患者纳入了 3787 个因素，包括人口统计学、病史、检查、生物标志物和药物。使用已发表的框架，我们通过九种无监督 ML 方法（K-均值、K-均值++、K-模式、K-原型、小批量、凝聚层次聚类、Birch、高斯混合模型和共识聚类）识别了亚型。我们使用聚类内分布和可解释的人工智能（AI）技术对聚类进行了特征描述。我们评估了亚型在以下方面的表现：（1）内部有效性（数据集内；跨方法）；（2）预后有效性（预测 5 年全因死亡率、住院率和新发慢性病）；和（3）药物负担。
我们发现了四个 T2D 亚型：代谢型、早发型、晚发型和心脏代谢型。亚型的预测具有很高的准确性（F1 评分>0.98）。5 年全因死亡率、住院率、新发慢性病发生率和药物负担在 T2D 亚型之间存在差异。与代谢亚型相比，新发 T2D 患者的 5 年死亡率和住院率在晚发型亚型中最高（HR 1.95，1.85-2.05 和 1.66，1.58-1.75），在早发型亚型中最低（1.18，1.11-1.27 和 0.85，0.80-0.90）。慢性病的发生率在晚发型亚型中最高，在早发型亚型中最低。与代谢亚型相比，在调整年龄、性别和 T2D 前药物后，晚发型亚型（1.31，1.28-1.35）和早发型亚型（0.83，0.81-0.85）在 T2D 发病后 5 年内最有可能和最不可能分别接受药物治疗。
在迄今为止最大的基于 ML 的新发 T2D 研究中，我们发现了四个不同的亚型，这可能对病因学、治疗学和风险预测具有潜在的未来意义。
