Artificial Intelligence (AI) models have huge potential to predict disease progression through analysis of health records, facilitating a more personalised healthcare service. This predictive capability is crucial in enabling more proactive health management of patients with chronic or acute illnesses related to lifestyle, genetics and socio-economic factors. Despite the existence of various predictive algorithms for certain diseases, a gap exists in comprehensive models that predict a wide range of medical conditions. Transformer models take inspiration from language models, promising to address this gap by modelling intricate temporal dependencies in health-related data.
Scientists from multiple institutions have now developed Delphi-2M, a sophisticated AI model that uses the GPT structure to anticipate disease progression in large populations. Using 400,000 participants’ data from the UK Biobank, it predicts over 1,000 diseases and deaths. The model analyses prior health records, demographic data and lifestyle factors, providing comprehensive future health trajectories for individuals. It also offers insights into groups of diseases and their time-dependent influences. After validation against 1.9 million Danish health records, the Delphi-2M model shows accurate modelling of population health and how past events impact future health outcomes, thus demonstrating its powerful predictive potential.
Delphi-2M significantly aligns its predictions of over 1,000 diseases with observed age and sex trends. It effectively models different disease patterns, such as early-life chickenpox peaks and increases in other conditions related to age. The model’s predictions are updated with new data and exhibit substantial variation between individuals for diseases like septicemia. With AUCs averaging 0.8, Delphi-2M’s performance equals that of established risk models like the Framingham model for cardiovascular disease. Confirmation of the model’s reliability in predicting short and long-term disease trajectories through calibration and longitudinal validation with the UK Biobank data demonstrates its effectiveness in delivering broad range disease predictions.
Generative models like Delphi-2M can predict future disease trajectories based on past health records. Upon evaluating 100,000 sampled trajectories, Delphi-2M reflects observed disease rates and incidences up to 70 years of age accurately. It consistently outperforms base age-sex models with an average accuracy of 17% in the first year, decreasing to 14% over 20 years. It effectively identifies high- and low-risk groups, predicting disease burdens spanning two decades. Additionally, Delphi-2M’s synthesized trajectories, which do not replicate training data, have practical uses, such as training new models, thereby helping preserve data privacy and broadening potential applications.
Delphi, a modified GPT-2 model, has been designed to forecast health trajectories by analysing sequences of primary ICD-10 diagnoses, supplemented with lifestyle data, such as sex, BMI, smoking, and alcohol usage. Training data from the UK Biobank and external validation with Danish health records were engaged. Delphi substitutes GPT-2’s discrete positional encoding with continuous age-based coding, introducing an additional feature to predict the time between health events. As a result, Delphi outperforms standard GPT models in predicting the timing and sequence of health events, disease onset, and progression.
Delphi-2M offers huge potential in healthcare planning, personalised medicine and understanding complex disease interactions. Future upgrades could incorporate more health data, such as genomics and wearables, making it even more powerful. However, while effective, it also inherits biases from its training data and must, therefore, be used with caution. I encourage you to check out the paper and code. All credit for the exciting research goes to the project’s researchers.
