Artificial Intelligence has significant potential to revolutionize healthcare by predicting disease progression using extensive health records, enabling personalized care. Multi-morbidity, the presence of multiple acute and chronic conditions in a patient, is an important factor in personalized healthcare. Traditional prediction algorithms often focus on specific diseases, but there is a need for comprehensive models that can predict a range of conditions. Transformer models, such as those inspired by LLMs (Large Language Models), show promise in overcoming these challenges by modeling complex time dependencies in health data.
Researchers have developed an AI model, Delphi-2M, based on the GPT (Generative Pre-training) architecture. Delphi-2M leverages data from 400,000 UK Biobank participants to predict over 1,000 diseases and deaths. It does this by analyzing past health records, demographics, and lifestyle factors. This model generates future health trajectories for individuals and provides insights into disease clusters and their impacts over time. The model was validated against 1.9 million Danish records and demonstrated its ability to model population health and reveal how past events affect future health outcomes.
The accuracy of Delphi-2M in predicting over 1,000 diseases aligns closely with observed age and sex trends. It models various disease patterns effectively and is continuously updated with new data. Its performance rivals established risk models. The model exceeded basic age-sex models with its average accuracy of 17% in the first year, dropping to 14% over 20 years. It identifies high- and low-risk groups and accurately predicts disease burdens over two decades. The synthetic trajectories generated by Delphi-2M, which do not duplicate training data, have real-world uses, like training new models, thus ensuring data privacy.
The model is a modified GPT-2 designed to predict health trajectories by analyzing sequences of top-level ICD-10 diagnoses, supplemented with lifestyle data. Delphi-2M uses an additional head to predict the time between events, allowing it to accurately model the timing and sequence of health events, which surpasses standard GPT models in predicting disease onset and progression.
Despite being effective, Delphi-2M may inherit biases from its training data and must be used cautiously. However, its flexible architecture allows for the integration of additional health data. This makes it a potentially valuable tool for healthcare planning, personalized medicine, and understanding complex disease interactions. The research and development of Delphi-2M highlights the potential of AI in revolutionizing healthcare and paves the way for more comprehensive and personalized patient care.
