Clarós Alejandro, Ciudin Andreea, Muria Jordi, Llull Lluis, Mola Jose Àngel, Pons Martí, Castán Javier, Cruz Juan Carlos, Simó Rafael
Higia.ai, Barcelona, Spain.
Diabetes and Metabolism Research Group, VHIR, Endocrinology Department, Vall d'Hebron University Hospital, Autonomous University Barcelona, Barcelona, Spain.
Eur J Public Health. 2025 Aug 1;35(4):642-649. doi: 10.1093/eurpub/ckaf098.
Metabolic syndrome (MetS) is related to non-communicable diseases (NCDs) such as type 2 diabetes (T2D), metabolic-associated steatotic liver disease (MASLD), atherogenic dyslipidaemia (ATD), and chronic kidney disease (CKD). The absence of reliable tools for early diagnosis and risk stratification leads to delayed detection, preventable hospitalizations, and increased healthcare costs. This study evaluates the impact of Transformer-based artificial intelligence (AI) model in predicting and managing MetS-related NCDs compared to classical machine learning models. Electronical medical data registered in the MIMIC-IV v2.2database from 183 958 patients with at least two recorded medical visits were analysed. A two-stage AI approach was implemented: (1) pretraining on 60% of the dataset to capture disease progression patterns, and (2) fine-tuning on the remaining 40% for disease-specific predictions. Transformer-based models was compared with traditional machine learning approaches (Random Forest and Linear Support Vector Classifier [SVC]), evaluating predictive performance through AUC and F1-score. The Transformer-based model significantly outperformed classical models, achieving higher AUC values across all diseases. It also identified a substantial number of undiagnosed cases compared to documented diagnoses fold increase for CKD 2.58, T2D 0.78, dyslipidaemia 1.89, hypertension 3.33, MASLD 5.78, and obesity 4.07. Diagnosis delays ranged from 90 to 500 days, with 35% of missed intervention opportunities occurring within the first five appointments. These delays correlated with an 84% increase in hospitalizations and a 69% rise in medical procedures. This study demonstrates that Transformer-based AI models offer superior predictive accuracy over traditional methods by capturing complex temporal disease patterns. Their integration into clinical workflows and public health strategies could enable scalable, proactive MetS management, reducing undiagnosed cases, optimizing resource allocation, and improving population health outcomes.
代谢综合征（MetS）与2型糖尿病（T2D）、代谢相关脂肪性肝病（MASLD）、致动脉粥样硬化性血脂异常（ATD）和慢性肾脏病（CKD）等非传染性疾病（NCDs）相关。缺乏用于早期诊断和风险分层的可靠工具会导致检测延迟、可预防的住院治疗以及医疗成本增加。本研究评估了基于Transformer的人工智能（AI）模型与传统机器学习模型相比，在预测和管理与MetS相关的非传染性疾病方面的影响。分析了MIMIC-IV v2.2数据库中登记的183958名至少有两次记录就诊的患者的电子医疗数据。实施了两阶段AI方法：（1）在60%的数据集上进行预训练以捕捉疾病进展模式，（2）在其余40%的数据上进行微调以进行疾病特异性预测。将基于Transformer的模型与传统机器学习方法（随机森林和线性支持向量分类器[SVC]）进行比较，通过AUC和F1分数评估预测性能。基于Transformer的模型显著优于传统模型，在所有疾病中均实现了更高的AUC值。与已记录的诊断相比，它还识别出大量未诊断病例，CKD增加2.58倍、T2D增加0.78倍、血脂异常增加1.89倍、高血压增加3.33倍、MASLD增加5.78倍、肥胖增加4.07倍。诊断延迟为90至500天，35%的干预机会错失发生在前五次预约期间。这些延迟与住院率增加84%和医疗程序增加69%相关。本研究表明，基于Transformer的AI模型通过捕捉复杂的疾病时间模式，比传统方法具有更高的预测准确性。将它们整合到临床工作流程和公共卫生策略中，可以实现可扩展的、积极主动的MetS管理，减少未诊断病例，优化资源分配，并改善人群健康结果。
