51吃瓜

There is a growing potential for AI to revolutionise healthcare as the technology evolves. Large language models (LLMs) can process significant amounts of health data in a fraction of the time that current methods take, paving the way for more efficient diagnoses and interventions for patients. 

Advances in biomedical AI enable processing of multimodal data, such as a mix of scans, genomic data and patients’ electronic health records, to provide more accurate prognoses. “Deep learning and advanced machine learning models are particularly effective at fusing these heterogeneous modalities, allowing for richer insights that traditional statistical models often miss,” says Tanujit Chakraborty, associate professor of mathematics and statistics at Sorbonne University Abu Dhabi. “These models excel at capturing the complex temporal and spatial dependencies in healthcare data, which are critical for tracking disease progression and forecasting outcomes.” 

Recent advancements, especially in graph-based and temporal neural architectures, will be transformative in enhancing the accuracy of predictions, patient monitoring and early intervention, Chakraborty says.

Samuel Feng, assistant professor of mathematics and statistics at the university, says that the interpretability of AI models is crucial for achieving this. “When an AI does something, it’s important for the person using it to understand why. For example, if an LLM suggests that we operate on someone, that would only be viable if the doctor using it can vouch for it and is part of the decision process,” he says. 

While some medical practitioners are beginning to use biomedical AI for low-stakes tasks, such as detecting type 2 diabetes and supporting patients in making lifestyle changes, other applications depend on ensuring that the output can be interpreted effectively. Feng’s research focuses on such interpretable models, which act like a scoring system to assign whether a data point is more or less likely to be correct. This helps the model to also communicate the uncertainty in its response, enabling doctors and patients to make better risk-calibrated decisions. 

Chakraborty’s research involves machine learning models for analysing temporal and spatiotemporal data in healthcare. “This model has shown strong performance in early outbreak detection and epidemic case prediction, making it a practical tool for public health planning and timely intervention,” he says. “We created a reinforcement learning framework tailored for zero-inflated count data, which is particularly relevant in mobile health applications such as apps and wearable devices.” His recent projects focus on Bayesian machine learning methods for dynamic treatment, which can help to design personalised and evolving treatment regimens for patients. 

Discussions around ethical issues in biomedical AI reflect what’s happening in the AI sphere more broadly, says Feng. A significant challenge is AI literacy, he says. “If you use the analogy of a car, most people can look at a car and know how it works. But there isn’t such an understanding of LLMs – they are a complete black box. This lack of literacy is dangerous.”

about Sorbonne University Abu Dhabi.