A new artificial intelligence (AI) technology can estimate a person’s likelihood of getting more than 1,000 diseases, in some cases delivering a prognosis decades in advance.
The Delphi-2M model estimates a person’s risk of developing illnesses including cancer, skin disorders, and immunological disorders up to 20 years in advance based on lifestyle variables and medical data. Despite being trained on a single UK data set, Delphi-2M’s multi-disease modeling may one day assist physicians in identifying high-risk individuals, enabling the early implementation of preventative interventions. A research that was published in Nature today describes the model.
The tool’s capacity to simulate several illnesses simultaneously is “astonishing,” according to Stefan Feuerriegel, a computer scientist who has created AI models for medical applications at the Ludwig Maximilian University of Munich in Germany. He claims that it is capable of producing whole future health trajectories.
Oracle of Health
Researchers have already created AI-based tools that forecast an individual’s chance of contracting specific diseases, such as cardiovascular disease and various types of cancer. However, according to data scientist Moritz Gerstung of the German Cancer Research Center in Heidelberg, who co-authored the paper, the majority of these tools only calculate the risk of one illness. “To provide a thorough response, a health care provider would need to run dozens of them,” he claims.
A generative pre-trained transformer (GPT), a sort of large language model (LLM) that serves as the foundation for AI chatbots like ChatGPT, was updated by Gerstung and his colleagues in order to overcome this issue. In response to a query, GPTs produce results that are statistically likely based on their extensive training data.
The modified LLM was created by the authors to predict a person’s risk of acquiring 1,258 diseases based on their prior medical records. Age, sex, body mass index, and health-related behaviors like drinking alcohol and smoking cigarettes are also taken into account by the model. Researchers used data from 400,000 individuals in the UK Biobank, a long-term biomedical monitoring program, to train Delphi-2M.
For the majority of diseases, Delphi-2M’s forecasts were as accurate as or better than those of existing models that calculate the likelihood of contracting an illness. Additionally, it outperformed a machine-learning system that predicts the risk of a number of diseases based on biomarkers, which are amounts of particular chemicals or substances in the body. “It was incredibly effective,” Gerstung adds.
When it came to predicting the future course of diseases with known development patterns, like some forms of cancer, Delphi-2M performed best. For a period of up to 2 decades, the model estimated the likelihood that an individual might develop each disease based on the data in their medical records.
Early-warning system
In the Danish National Patient Registry, a nationwide database that has monitored hospital admissions for over 50 years, Gerstung and his associates evaluated Delphi-2M using health data from 1.9 million individuals. According to the authors, the model’s predictions for registry members were only marginally less accurate than those for UK Biobank participants. This shows that when the model is used with data sets from national health systems other than the one it was trained on, it can still produce quite accurate predictions, according to Gerstung.
“Delphi-2M is an intriguing addition to the emerging field of modelling multiple diseases simultaneously, but it has limitations,” says Degui Zhi, a bioinformatics researcher at the University of Texas Health Science Center at Houston who creates AI models. The UK Biobank data, for example, only recorded individuals’ initial encounters with an illness. A person’s history of disease is “important for the modelling of personal health trajectories,” according to Zhi.
To increase Delphi-2M’s reach, Gerstung and his associates will assess the accuracy of the model using data sets from numerous countries. According to him, it will be crucial to consider how this data may be integrated to create algorithms that are even more accurate.
