Healthcare (Commonwealth Union) – A recent artificial intelligence system that can generate virtual patient profiles and forecasting personalised health outcomes is being described as a potential breakthrough for clinical trials.
An Australian research team at the University of Melbourne trained an existing large language model (LLM) using three major datasets containing thousands of electronic medical records.
Their AI model, named DT-GPT, examined information from patients with Alzheimer’s disease, non-small cell lung cancer, and those receiving intensive care.
By producing digital replicas of these patients, the model projected how their health might evolve over time with treatment, offering insights into the likely progression of their conditions.
DT-GPT achieved its accurate forecasts by drawing on its built-in understanding of medical literature and analysing each patient’s medical background, including lab data, diagnoses, and treatments.
The model did not receive any data about patients’ eventual health outcomes, which enabled the researchers to test the accuracy of its forecasts.
According to lead researcher Associate Professor Michael Menden, they generated a digital twin for every patient by commencing the model with their unique clinical characteristics.
He indicated that for an example, they formed virtual twins of 35,131 intensive care unit (ICU) patients and with accuracy forecasted what would take place in their magnesium levels, oxygen saturation and their respiratory rate in a period of 24 hours, based on their laboratory findings from the prior day.
Overall, DT-GPT delivered better predictive performance than 14 leading machine learning systems tested alongside it.
According to the researchers, the model may eventually help forecast clinical trial results, speeding up drug development while reducing costs and improving efficiency.
The period of trials which at times can take up to 20 years, often means that many individuals who may desperately requires new life saving treatments often miss out. A procedure that can harness the power of AI could become a gamechanger in the world of healthcare.
Associate Professor Menden indicated that this innovation marks a move from reactive care toward proactive, personalised medicine.
He further pointed out that it may permit clinicians to foresee when a patient’s condition is likely to worsen, giving them the chance to step in sooner.
“It could also be used to predict negative side effects of medications, allowing doctors to tailor treatment plans to suit each patient’s unique characteristics and medical history, ultimately increasing the chances of a positive health outcome.”
Researchers of the study indicated that the model can rapidly make sense of complex and unstructured information, and it features a chatbot-style interface that lets users ask questions and explore how it arrives at its predictions.
Because DT-GPT uses generative AI, it’s also capable of making zero-shot predictions—informed estimates about clinical measurements it was never specifically trained on.
Associate Professor Menden explained gave an example indicating that it is similar to asking the system to guess how tall someone will become without giving any height records, relying only on their past weight and shoe size.
He further indicated that their model precisely predicted the way lactate dehydrogenase (LDH) levels were altered in non-small cell lung cancer patients 13 weeks after they commenced the therapy, with no training of the model for this purpose.
“We compared it to traditional machine learning models, which were specifically trained for 69 clinical variables, including LDH, which we in comparison only educated guessed.
“Very surprisingly, the DT-GPT’s zero-shot predictions, its untrained guesses, were more accurate in 18 percent of cases.”
The study was recently featured in NPJ Digital Medicine.
The group behind the DT-GPT artificial intelligence system, working alongside the Royal Melbourne Women’s Hospital, has since launched a new venture aimed at creating digital twin models for people with endometriosis—demonstrating how broadly this technology can be applied.
