New AI Tool Predicts Patient Health with Digital Twins

A groundbreaking artificial intelligence tool, known as DT-GPT, has been developed to create virtual representations of patients, enabling precise predictions of individual health trajectories. Researchers at the University of Melbourne utilized extensive datasets comprising thousands of electronic health records to enhance this innovative model. The tool has been recognized as a potential transformative element for clinical trials.

DT-GPT was trained using data from patients diagnosed with either Alzheimer’s disease, non-small cell lung cancer, or those admitted to intensive care units. By analyzing comprehensive medical histories, including laboratory results, diagnoses, and treatments, the AI tool generates digital twins of patients, forecasting how their health may evolve over time during treatment.

Associate Professor Michael Menden, the lead researcher, explained the process: “For each patient, we created a virtual replica by initializing the model with their individual clinical profile.” In one instance, the model successfully predicted critical health indicators—such as magnesium levels, oxygen saturation, and respiratory rates—of 35,131 ICU patients over a 24-hour period, based on their previous day’s laboratory results.

The DT-GPT model demonstrated superior predictive accuracy compared to 14 other advanced machine learning models. Its potential application in simulating clinical trial outcomes may significantly expedite drug development, making it both faster and more cost-effective. “This technology paves the way for a shift from reactive to predictive and personalized medicine,” Menden noted.

Advancements in Predictive Medicine

The innovation allows healthcare professionals to anticipate potential health declines, facilitating earlier intervention. Furthermore, the model can predict adverse medication side effects, enabling healthcare providers to customize treatment plans based on each patient’s unique medical history. This capability could ultimately enhance the likelihood of achieving favorable health outcomes.

DT-GPT excels in interpreting complex and disordered data. It features a conversational interface, allowing users to engage with the model similar to a chatbot, gaining insights into the reasoning behind its predictions. The tool employs generative AI to provide “zero-shot predictions,” which are educated guesses regarding laboratory values without prior training on specific data.

To illustrate this capability, Menden provided an analogy: “It’s like asking the model to predict how tall someone will grow without providing the person’s height records and only giving their previous weight and shoe sizes.” Remarkably, DT-GPT accurately forecasted changes in lactate dehydrogenase (LDH) levels in patients with non-small cell lung cancer 13 weeks into their therapy, despite not being expressly trained for this prediction.

In a comparative analysis against conventional machine learning models that focused on 69 clinical variables, including LDH, DT-GPT’s untrained guesses outperformed the traditional models in 18% of cases.

The findings from this research are detailed in the journal npj Digital Medicine, with additional contributions from Nikita Makarov and colleagues. As the healthcare sector continues to seek innovative solutions, DT-GPT stands at the forefront, promising to reshape the landscape of personalized medicine and clinical trials.

This advancement not only enhances understanding of patient health but also signifies a significant step towards integrating AI technology into everyday medical practices, ultimately aiming to improve patient care worldwide.