Скачать или смотреть Delphi-2M: Disease Progression Modeling with Generative Transformers

Details the development and evaluation of Delphi-2M, a novel generative transformer model designed to predict the progression of human disease. This model adapts Large Language Model (LLM) architecture to interpret a patient's health history as a structured sequence of events, effectively learning the "grammar" of disease progression and multi-morbidity. Key architectural innovations include continuous age encoding and a dual output head that predicts both the next disease and the time until it occurs, transforming the model into a quantitative epidemiological forecasting tool. Trained on data from the UK Biobank and externally validated on Danish registries, Delphi-2M demonstrates strong predictive power across over a thousand diseases, achieving performance comparable or superior to specialized clinical risk scores for endpoints like death and dementia. Furthermore, its generative capability allows for the creation of privacy-preserving synthetic health data, and Explainable AI (XAI) techniques confirm that the model learns clinically meaningful relationships, despite exhibiting some biases related to data collection artifacts.