How Tal Arbel is building trustworthy AI to predict and treat neurological diseases like multiple sclerosis

Imagine a future where we don’t just predict the onset of chronic conditions like multiple sclerosis (MS), but map their entire clinical course with precision. By forecasting how specific interventions will reshape an individual’s unique trajectory, we can move beyond generalized care and toward truly personalized medicine.

These are the questions driving the work of Tal Arbel, a Canada CIFAR AI Chair at Mila and a Professor at McGill University. Arbel is at the forefront of building equitable, safe and trustworthy AI systems designed for the high-stakes complexities of real-world clinical deployment. Her research sits at a critical intersection: where computer vision and probabilistic machine learning meet to navigate real-world complexities in medical imaging analysis.

Identifying the hidden patterns 

In neurological diseases like MS, the stakes for precision are incredibly high. These conditions evolve in highly heterogeneous and subtle ways. For many diseases, “minute changes in brain structure or lesion patterns can profoundly alter a patient’s physical and cognitive abilities in ways that are not fully understood,” Arbel explains.

Traditional radiology struggles with these nuances because the patterns are too subtle and multidimensional for manual detection. Arbel notes that analyzing static clinical data, a single snapshot in time, is insufficient for chronic conditions that progress over time. To truly forecast a patient’s future, AI must be able to model complex temporal dynamics and act on data that changes over time.

Building 3D medical foundation models

While modern AI has achieved breakthroughs in 2D image analysis, current models are not well-equipped to handle the complex 3D structure of brain scans. To bridge this gap, Arbel’s team is building specialized 3D medical foundation models. This will permit models to integrate multi-modal data, combining high-dimensional medical images with patient demographics and clinical history to recognize imaging signatures that predict whether a patient will progress rapidly or slowly.

This represents an opportunity to significantly advance the field of neurological care by moving from descriptive AI, which aims to identify a tumor, to prescriptive AI, which can forecast disease progression. 

PRISM: simulating healthy images

One of Arbel’s most innovative breakthroughs is PRISM: High-resolution & Precise Counterfactual Medical Image Generation using Language-guided Stable Diffusion. PRISM is a generative AI model that allows clinicians to ask “what if” questions.

By analyzing a patient’s MRI and generating a hypothetical healthy image for that individual, researchers can precisely locate the physical structures responsible for the patient’s symptoms.

Crucially, PRISM acts as a safety check against biases inherent in medical imaging data. It can reveal spurious correlations within datasets—for example, when medical devices or image artifacts frequently co-occur with a disease rather than the actual pathology. By exposing these underlying data flaws and generating realistic counterfactual images to balance the dataset, PRISM helps make downstream AI tools fairer and better equipped to handle diverse populations.

A new framework for treatment

To move toward truly personalized medicine, Arbel’s team developed the first stochastic causal temporal framework. By leveraging causal inference, a field of AI that identifies specific cause-and-effect relationships, the model captures the continuous evolution of disease. It applies Neural Stochastic Differential Equations (NSDE) to a patient’s high-dimensional images (MRI) and tabular data, allowing clinical teams to predict both factual and counterfactual trajectories.

The team validated this framework using data from MS patients across six major clinical trials, accurately predicting clinical disability scores up to two years in advance. This is significant because, while many clinical trials fail because the treatment doesn’t work for the average patient, this model looks more deeply. By accounting for the heterogeneity, or biological diversity, within a group, it can pinpoint specific clusters of patients who responded positively, thus allowing the team to identify effective treatments that may have been discounted because they didn’t work for most patient groups. 

“This suggests a path forward for precision medicine in MS,” says Arbel. “Rather than treating all patients with the same drug, multimodal imaging-based AI has the potential to match them to their optimal treatment.”

Looking ahead to the future and the Canadian ecosystem

Looking ahead, Arbel is exploring agentic AI, systems that can plan, take action, and learn from feedback, mirroring the way clinicians think.

She credits the Canada CIFAR AI Chairs program with providing the academic freedom and agility required to pivot into high-promise areas such as causal inference and multimodal language models (advanced AI systems that can process multiple types of data, such as 3D imaging data and clinical data).

“The Canadian ecosystem is the glue that holds our AI community together,” Arbel says. “It proves to industry—including pharmaceutical companies and clinical collaborators—that Canada is fully equipped to translate theoretical AI into trustworthy tools that directly improve patient care.”