AI in the mainstream has become synonymous with LLMs, giving students an easy way out of assignments or a tool to generate content. But is there a way in which AI has a tangible benefit? European scientists may have an answer.
A team at the European Molecular Biology Laboratory (EMBL) in Cambridge and the German Cancer Research Center introduced a new AI model for healthcare called Delphi-2M, which can predict more than 1,000 conditions a person might face in the future. Its creators hope that it could predict conditions like Alzheimer’s disease or cancer, which affects millions of people each year.
Authors have taken inspiration from large language models, such as Gemini, which are trained on enormous amounts of text scraped from the internet. These models learn to select the word most likely to come next in any given sentence. Similarly, Delphi-2M AI model analyses data from 400,000 anonymous participants to predict healthcare conditions.
The difference from typical LLMs lies in its ability to account for the time between conditions and the patients’ life events. Creating this feature didn’t come without problems, as early version sometimes predicted diagnoses for people who had already died.
The model was subsequently tested on data from 1.9m Danes, yielding varying results. Events that follow from a specific condition, like diabetes, had more accurate predictions, while more random external factors, like a virus, were harder to predict.
It may take up to ten years before we will see healthcare Gen AI used in daily healthcare checkups. Nonetheless, the model has proven valuable for research, as clustering conditions allows exploration of relationship between diseases. AI is already present in hospitals, mostly assisting with analysing healthcare data. A well-known example, serving over 300,000 patients annually, is Powerful Medical, start-up that interprets electrocardiograms enabling early diagnosis of cardiovascular conditions.
However, there are downsides. A series of recent studies reported that AI models across the healthcare sector led to biased results for women and ethnic minorities. The problem lies in the datasets used for training, content from the internet, which existing societal biases reflected in the LLMs responses. Researchers from MIT have suggested that one way to reduce bias in AI is to filter which data should be used for training.
References:
Heikkilä, M. (2025, September 19). AI medical tools downplay symptoms in women and ethnic minorities. Subscribe to read. https://www.ft.com/content/128ee880-acdb-42fb-8bc0-ea9b71ca11a8
Guardian News and Media. (2025, September 17). New AI tool can predict a person’s risk of more than 1,000 diseases, say experts. The Guardian. https://www.theguardian.com/science/2025/sep/17/new-ai-tool-can-predict-a-persons-risk-of-more-than-1000-diseases-say-experts
A new AI model can forecast a person’s risk of diseases across their life. (n.d.). https://www.economist.com/science-and-technology/2025/09/17/a-new-ai-model-can-forecast-a-persons-risk-of-diseases-across-their-life
Really interesting post Simon. As AI is improving I think it becomes almost a moral duty to incorporate this technology in the medical sector, because it has the potential to help save so many lives. But having AI as your own GP is something that doesn’t sit right with me yet.
I think AI can be incredibly helpful at giving triage advice and simple suggestions about how to temporarily tone some symptoms down, but I struggle with imagining AI becoming a proper GP. Health advice and diagnosis is very risky, as a wrong diagnosis can really deteriorate health, so this is why we have properly trained GPs that are ready to take responsibility for their actions. A simple mistake in the medical sector can ruin someone’s life, and who would take responsibility if the AI model makes a wrong diagnosis? The engineers of the AI model, the person that decided which data to use to feed the model, or the directors of the hospital that decided to implement the AI model? You could argue for any one of them, so when accountability becomes blurry, I dont think good things can result from it.
However, I do agree with you that AI should be implemented in the medical sector, but they should be used merely as a tool, alongside a trained doctor.