Photo courtesy of the University of Melbourne
Researchers from the University of Melbourne have created a generative AI tool that predicts patients' disease progression more accurately than existing machine learning-based models.
Dubbed DT-GPT, the AI model can create digital twins of patients and forecast how their health would progress over time under treatment.
WHAT IT'S ABOUT
The model is based on the open-source biomedical LLM, BioMistral 7B DARE, which the research team further trained using electronic patient health record databases from three different patient cohorts – patients in the intensive care unit, those with Alzheimer’s disease, and non-small cell lung cancer (NSCLC) patients.
It is fed with an individual patient's clinical profile, including laboratory results, diagnoses, and treatments, which it then uses to create their virtual replica or digital twin.
In a media release, the university described the model as having a conversational interface and the ability to quickly interpret dense and messy data.
Lead researcher and associate professor Michael Menden said DT-GPT can also make zero-shot predictions, or educated guesses about laboratory values.
For example, DT-GPT accurately predicted how lactate dehydrogenase levels in NSCLC patients 13 weeks after starting therapy without being trained for that purpose. The team then compared it with trained ML models. "Very surprisingly, the DT-GPT's zero-shot predictions – its untrained guesses – were more accurate in 18% of cases," A/Prof Menden noted.
Based on findings published in NPJ Digital Medicine, the DT-GPT model outperformed 14 other forecasting models, including classical statistical models, time-series ML, and deep learning or time-series neural architectures, across the three EHR datasets.
WHY IT MATTERS
UniMelb says DT-GPT may assist with predictive and personalised medicine.
"It could enable doctors to anticipate if their patient’s health will deteriorate so they can intervene earlier," A/Prof Menden said.
"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," he added.
The model can also be used to simulate clinical trial outcomes, potentially driving "faster, cheaper, and more efficient" drug development, the university noted.
The DT-GPT team is now collaborating with the Royal Melbourne Women's Hospital to set up a new company that develops digital twins for endometriosis patients.
