In the world of prevention and personalized medicine, the ability to anticipate with a long time in advance is key, even decades, the possibility that a disease appears, to stop the processes that would produce it long before it started harming. That is what seems to be achieved thanks to artificial intelligence, according to a study published today in the magazine Nature.
Throughout life, we start being well more frequently, from time to time we spend one day sick, and over time the chronic ailments begin to appear. These patterns affect each individual in a different way, depending on the inheritance, lifestyle or their socioeconomic status. To understand a person’s health well and value the risks that stalk it in the future, it is not enough to take the isolated diagnoses that he received throughout his life; It is necessary to understand the evolution of each person, know the diseases they suffered to know how to influence each other and promote specific life changes or recommend diagnostic tests that monitor concrete ailments with the most likely to appear.
Today, a group of researchers from the European Institute of Bioinformatics, the DKFZ (German Cancer Research Center) and several Danish institutions proposes to apply the same technology that gives life to the great language models – such as Chatgpt – to learn and predict the natural history of more than a thousand diseases at the same time. The resulting model, baptized as Delphi-2M, is capable of identifying disease patterns from medical records, lifestyle factors and previous health conditions.
“The most unexpected finding was that the model can predict more than 1,000 diseases. We would have expected that it would work with some, but that it would fail with many others. This shows how interconnected many diseases are and highlights the need to investigate the underlying mechanisms that connect them,” explains about their results Moritz Gerstung, director of the artificial intelligence division in Oncology of the DKFZ study.
The algorithm has been trained with data of 400,000 people from the United Kingdom and validated with records of almost two million patients in Denmark, and is able to project health trajectories, both at the population and individual level, of up to two decades.
As with weather predictions, this model does not offer certainties, but probabilities. More than guessing exactly what will happen to a specific person at a given time, calculates the chances of suffering certain diseases in a specific period. As with time, short -term predictions are more reliable than those that try to predict a farther future. When they are calculated if someone will suffer a heart attack in the next 10 years, the model is right about seven out of ten cases. When the temporal period is extended to two decades, it stays at 14%, somewhat greater than 12% that is achieved knowing age and sex.
Continuing with the case of the infarction, according to the model, in the cohort of the Biobanco of the United Kingdom, men between 60 and 65 can have an annual risk of 4 out of 10,000 to 1 in 100, depending on their medical background and their life habits. In women the average risk is lower, but the probabilities dispersion is similar. The most relevant thing is that, when comparing the predictions of the model with real biobanco data that were not used in training, it was found that the calculated risks coincided with the observed incidence of cases in different age and sex groups. This shows that estimates faithfully reflect real population trends.
Delphi-2M reaches a precision comparable to the best specific models for diseases such as dementia or myocardial infarction, and surpasses mortality prediction algorithms. Only in the case of diabetes, a blood test marker (HBA1C glycosylated hemoglobin) remains more reliable. In addition, the study identified diseases that increase the risk of other, such as mental disorders or some tumors of the female reproductive system.
On the possibility that knowing so in advance of diseases that they are only a possibility makes us all preventive patients, Gerstung believes that more studies are needed to raise how this knowledge can benefit patients. That would require that the possible AI applications as a medicine assistant “should be tested in randomized clinical trials, in which a group receives medical visits with support from AI and another group without it. After a period of monitoring, it will be evaluated if the group assisted by IA obtained greater benefits compared to traditional consultations,” he says. “This can also include subjective evaluations of people’s well -being to assess the emotional effects of knowing or not their risks,” he concludes.
In the section on possible risks of such a powerful tool for health prediction, such as discrimination by the insurers of patients with risks that make them little interesting such as clients, Guillermo Lazcoz, a member of the Ethics Committee of the Research of the Carlos III Health Institute, considers that the application of AI to the processing of large health databases adds “one more layer of risks to which we already knew what we already knew” The hands of a bank to use them to know, before granting a loan, if the client is prone to contracting a type of cancer or having a heart attack.
“The AI can identify a person from data that were supposed to anonymous, which demands new protection measures,” says Lazco. To apply these measures, “in Europe safe spaces for data processing are being developed, where the data does not travel and access to third parties is already limited in time and a purpose,” he explains. Finally, he warns that it is not the same to talk about organizations such as the Biobanco of the United Kingdom, used in the study today Naturewhich has strict controls, that from companies like 23 Andme, in which one can analyze their DNA to know its lineage and that has already been involved in scandals due to problems with data protection.
Mikel Recuero, researcher at the University of the Basque Country (EHU) and lawyer specialized in data protection, agrees that, at least in the European sphere, there are many layers of control that seek to prevent the improper use of biomedical data. “Access to Biobancos already implies a first ethical filter, because researchers must justify the scientific purpose of their application and cannot use samples for spurious purposes,” he says. “To this are added data protection controls: if the information is identifiable, the regulations forces to restrict its use to authorized purposes, avoiding, for example, applications in insurance or banking,” he adds.
In this sense, “the new regulation of the European Health Data Area reinforces this logic by expressly prohibiting commercial decisions – as the modification of premiums of insurance – based on genetic information,” he says. “Although risks never disappear at all, there are successive (ethical, regulatory, legal) mechanisms that act preventively, limiting the possibilities of discrimination and forcing to accredit a social benefit in each project that will use this data,” he concludes.
Models such as GPT-4 or Gemini learn language as a sequence of words. They predict the following word depending on the context and the researchers saw an analogy with health. The medical history of a person can also be understood as a sequence of events – diagnosis, risk factors, lifestyle habits – that follow a temporary order to make predictions.
At the moment, the model should be improved to be useful to take care of the health of real patients, but it is already a useful tool to better understand how diseases develop and how to progress, or evaluate the effects of lifestyle or past diseases influences the risk of future diseases.
One of the most innovative aspects of work is Delphi’s ability to generate synthetic health data. From partial information, the model can imagine complete trajectories that maintain the same statistical properties as real data, but without corresponding to any particular person. This protects the privacy of patients, since the data cannot be linked to real individuals and allow to train other AI models without accessing sensitive clinical data. This could, for example, calculate what can happen to the population’s health if obesity increases by 5%.
Now, there are already algorithms that predict the risk of suffering some diseases, such as heart problems or breast cancer, but that approach does not cover the real complexity of human health, in which, many times, multiple diseases that are conditioned between them coexist. In increasingly aging societies, the ability to provide for the burden of many diseases and design policies and investments will be critical to try to prevent them and be prepared when they arrive.
