A model has been created that anticipates the appearance of a thousand diseases 10 years from now.

BarcelonaOne of the fundamental tenets of medicine is to anticipate the onset of symptoms of any disease to improve the patient's prognosis. For example, starting treatment for someone at very high risk of suffering a heart attack can minimize the consequences or even prevent cardiovascular disease before it occurs. Or, if cancer is detected in very early stages, the patient will need less aggressive treatments, and their survival will be better than when the tumor is identified in more advanced stages or has metastasized and spread to other areas. Therefore, the scientific community is making a significant effort to find new early detection mechanisms that improve current ones. Now, an international team of researchers has developed a method to predict the risk and timing of onset of more than a thousand diseases and predict health outcomes more than ten years in advance.

"Medical events often follow predictable patterns," says Tom Fitzgerald, one of the study's authors and a researcher at EMBL's European Bioinformatics Institute (EMBL-EBI), who led this research together with the German Cancer Research Center (DKFZ). Following Fitzgerald's premise, the authors of this research, published this Wednesday in the journal Nature, They have trained an artificial intelligence model that, using medical records, is capable of estimating on a large scale how human health can change over time.

The international team conducted their research using data from more than 400,000 patients from the UK Biobank, an initiative that contains genetic and health data from half a million Britons and is one of the most powerful computer databases in Europe. Now, after testing the new technology, they have obtained very promising results. In fact, the model is "especially good" when the disease has clear and consistent progression patterns, such as some cancers, heart attacks, and sepsis, which is a very extreme reaction to a blood infection that can be fatal.

"By modeling how diseases develop over time, we can begin to explore when certain risks emerge and how to better plan early interventions. This is a major step towards more personalized and preventative approaches to healthcare," says Ewan Birney, Interim Executive Director of EMBL. However, it should be noted that the model they developed is "less reliable" at predicting the onset of diseases subject to more variables, such as mental health disorders or pregnancy complications, as these are often linked to unpredictable everyday events. However, the researchers emphasize that, like weather forecasts, this new AI model offers probabilities, not certainties, and is not yet ready for clinical use.

Chances of developing certain diseases

After training the AI model with data from patients, their diagnoses, and their lifestyles, the authors tested it with data from 1.9 million people. They clarify that it doesn't predict exactly what will happen to someone, but it does provide "well-calibrated" estimates of the likelihood of certain diseases or risk factors for developing them. They also emphasize that short-term predictions are more accurate than long-term ones. For example, the model predicts different levels of heart attack risk, and in the 60-65 age group, some men have a 4 in 10,000 chance per year, while others have approximately 1 in 100, depending on their previous diagnoses and their lifestyle (whether they smoke or don't exercise).

Furthermore, risks increase, on average, as people age. Despite the positive results, the authors warn that, like any AI model, it has limitations. For example, they used data primarily from people between the ages of 40 and 60 and admit that health events in childhood and adolescence are underrepresented. Furthermore, the model also contains "demographic biases" because the data they used does not include enough people from all ethnic groups. For all these reasons, the researchers insist that the model is not ready for clinical use. However, they believe that the results can already help understand how diseases develop and progress over time, explore how lifestyle and previous illnesses affect the long-term risk of developing others, and also simulate health outcomes with artificial patient data in situations where real data is difficult to obtain.

"This is the beginning of a new way of understanding human health and disease progression," says Moritz Gerstung, head of the AI in Oncology Division at the DKFZ. The expert believes that models like this one could help "personalize care and anticipate large-scale healthcare needs" in the future, and therefore argues that these tools need to be trained with more representative data. With increasingly aging populations and an increase in chronic diseases, Gerstung argues that anticipating citizens' healthcare needs will help healthcare systems plan better and allocate resources more efficiently, but first, more research and a robust regulatory framework for clinical use are needed.

Gustavo Sudre, professor of Genomic Neuroimaging and Artificial Intelligence at King's College London—who was not involved in the study—considers this research an important step toward making AI models in medicine scalable and interpretable. "It is encouraging to see that the model architecture has been deliberately designed to accommodate richer types of data, such as biomarkers, imaging, and even genomics," he highlights in statements to the Science Media Center (SMC). However, he believes this model is "well positioned to evolve into a truly multimodal precision medicine tool."