Researchers in the UK tested an AI tool called Foresight that creates digital twins of patients to forecast future health and treatment outcomes.
The idea of creating digital twins in various industries allows engineers to test systems in a simulation before deploying them in the physical world. AI tools like Foresight are now making this possible for healthcare practitioners.
Every time a patient visits a healthcare practitioner information is added to their electronic health record (EHR). Some of this data is structured (age, gender, ethnicity), but most of it is unstructured, like test results or notes a doctor might make.
Foresight uses a GPT-based model to transform this data into a model, or digital twin, of the patient. Because Foresight is trained on huge amounts of other patients’ EHR data, it is then able to forecast health outcomes like the kinds of diseases a patient is likely to develop, or their response to a certain kind of treatment.
James Teo, a professor at King’s College Hospital and study co-author, explained the significance of this. Teo said on X, “Unlike LLMs that simply predict the next word, Foresight forecasts possible futures for patients, representing possible multiverses to understand diseases.”
You could take a patient’s EHR and then simulate multiple versions of the patient to predict their health trajectory. Traditionally, a doctor would need to read a patient’s EHR, decide on a treatment option, and then evaluate the results after some time to monitor the effectiveness of the treatment.
With Foresight, a doctor could simulate multiple potential treatments with the model forecasting the short-term and long-term outcomes of each treatment. This is a far more cost-effective approach and spares the patient from the “Let’s try this,” approach many doctors have to resort to.
Results
The study, published in The Lancet Digital Health, explained how the researchers trained three different models of Foresight using hospital datasets from two UK hospitals and a publicly available dataset in the US, for a total of 811,336 patients.
Foresight was tasked with selecting the disorder a patient was most likely to develop from a list of 10 possible disorders. It accurately predicted the next disorder 68% and 76% of the time using the two UK datasets and 88% of the time when using US data.
When tasked with forecasting the next new biomedical ‘concept’, which could be a disorder, symptom, relapse, or medication, Foresight achieved a precision of 80%, 81%, and 91% respectively using the UK and US datasets.
The variance in performance shows how dependent AI tools are on having good-quality data.
As exciting as this application of AI is, the researchers note several challenges that need to be overcome. Finding ways to have the model properly weigh new treatments and interventions, or properly evaluate the importance of probability vs urgency and impact are just two examples.
The researchers are working on developing Foresight 2, which they say will be a more accurate model.
With new drug discovery and concepts like patient modeling, simulation, and forecasting, AI is set to have a significant impact on the quality of healthcare we get.
