Artificial Intelligence (AI) took center stage this month with the awarding of Nobel Prizes. The Nobel Prize in Physics was granted for foundational work in machine learning, while the Nobel Prize in Chemistry was given for using AI to predict protein structure. AI is also contributing significantly across the MS field. What are the current applications and limitations of AI? And what is its contribution in shaping the future of MS research and clinical care?
AI can offer a second opinion
AI models are trained on large datasets and learn recurring patterns or informative features. Based on these patterns, they make predictions about new, unseen data points. AI tools, especially machine learning and deep learning, can process and analyse complex, large clinical datasets, surpassing the limitations of human time, budget, and cognitive capacities [1]. They can assist humans in many tasks. In the healthcare system, making diagnoses of MS, prognoses and monitoring treatment may certainly benefit from the use of AI.
“AI is increasingly used in neuroradiology techniques to process images. It can speed up the process of acquiring scans of the brain and spinal cord”, Dr Arman Eshaghi, from the University College London, tells us.
Notably, AI holds promise to reduce the dose of gadolinium injection needed for contrast-enhanced brain scans, lowering health risks, costs, and scanning time [2].
Professor Douglas Arnold from the McGill University in Canada tells us, “Today, the focus is on the application of artificial intelligence in prognosis, particularly in predicting disability progression and disease activity – in terms of relapses and formation of new lesions. At the company where I work, we employ machine learning models for brain volume measurement and brain tissue segmentation in clinical trials to evaluate treatment efficacy. Quantitative MRI is also making its way into clinical practice. As for MS diagnosis, many labs are actively working on AI-based solutions. However, further studies with larger and more diverse populations are needed. Currently, AI does not outperform the average clinicians, who apply the McDonald criteria alongside their clinical expertise.”
By combining clinical and MRI data, machine learning algorithms can predict with reasonable accuracy which individuals with clinically isolated syndrome (CIS) will develop MS [3]. Machine learning has the potential to analyse electronic health records to better predict a person’s future health. By embedding millions of electronic health records into a biomedical knowledge graph known as SPOKE, researchers could create very detailed health profiles that look at both clinical and biological aspects. This method may enable the detection of the MS prodrome up to five years before an official diagnosis [4]. Additionally, machine learning models that learn clinical data collected over several years, have the potential to predict how the disease will progress [5].
Using AI to personalise treatments
Different treatments often work better for some patients than others. Researchers aim to improve clinical trials by focusing on individuals who are predicted to benefit the most from a treatment, a strategy known as predictive enrichment [6]. By not including people who are less likely to benefit from a trial, unnecessary risks can also be reduced. This approach could significantly decrease the number of participants needed, shorten the duration and costs of clinical trials – with results achievable in just one or two years – and help ensure that trials are completed. “This is an important issue especially to investigate treatments for progressive MS”, continues Professor Arnold. “The trial must be large, and therefore very expensive, and the drugs do not always succeed. We have recent examples of failed trials, where entire programs were dropped, with huge expenses and no benefit. This can make companies reluctant to invest in large trials. There is potentially a great benefit in being able to enrich a phase-II trial with people who are more likely to respond to a therapy, determining in short time with fewer individuals whether a drug has a potential to explore in a larger phase-III trial.”
Deep learning can help predict how individuals will respond to treatment based on clinical and demographic information, and baseline MRI metrics. With this method, scientists were able to estimate the individual effect of anti-CD20 therapies on disability progression [6].
Professor Tal Arbel from the McGill University explains, “Together with Professor Arnold, who serves as president of NeuroRx, we bring over 20 years of expertise in developing machine learning models for segmenting – that is, identifying and isolating – lesions in brain MRI scans of individuals with MS. Thanks to a grant from the International Progressive MS Alliance (IPMSA), we have access to MRI, clinical, and demographic data from 10,000 individuals with MS over time, as well as lesions labels and clinical outcomes. This has enabled us to develop modern deep learning models capable of detecting, segmenting, and counting lesions. We are interested in building models that integrate clinical, demographic data and medical images. These AI tools can offer clinicians much more personalised information about the individuals with MS. We would like to develop a tool that can predict future individual treatment response and outcomes under different treatments using baseline MRI and clinical information. We are working on a deep learning model that can predict new and enlarging lesions counts in follow-up MRIs under different treatments. By doing so, we will be able to estimate how an individual will respond to different therapies and predict lesion suppression under different treatment compared to placebo [7]. This has the potential to significantly improve clinical care and optimise clinical trial enrichment.”
Limitations and risks of AI
AI tools are highly promising and have the potential to substantially enhance MS research and clinical care in various ways. However, they are not without limitations. They can make mistakes and hide biases [1]. While they enable to produce more at a faster pace, we must ensure this does not lead to a diminished understanding [1].
Professor Arbel, who is also the director of the Probabilistic Vision Group and Medical Imaging Lab in the Centre for Intelligent Machines at the McGill University, emphasises the importance of integrating uncertainty into machine learning models. “In our lab, we focus on developing models that estimate uncertainties in their predictions for each treatment option and across different treatments. Importantly, uncertainties in predictions relate to their errors [8]. So, when the model says, `Hey, I am very confident`, it is more likely to be correct. In clinical trials, this model could help select participants for whom it has the highest confidence in predicting outcomes.”
Another challenge concerns the lack of transparency and interpretability that defines many machine learning models. Often referred to as “black boxes”, these models do not clearly explain their decision-making process [9]. “Clinicians really need to understand how the model arrived at its decision. We need to work on AI models that expose the imaging markers used to make their predictions”, Professor Arbel says.
Additionally, AI models can harbor and perpetuate hidden racial and gender biases, that often stem from the data they are fed [10]. “It is important to create fair and equitable models that identify and overcome biases in outcome predictions to ensure accuracy and generalisation across diverse clinical trials and groups of individuals with MS” says Professor Arbel. “These models have to work effectively not only for the majority of people, but for everyone, including underrepresented groups.”
At this stage, AI does not exceed the interpretations of the average clinical expert. Addressing these open challenges is essential for advancing the application of AI in MS research and clinical practice.
***
Written by Stefania de Vito
Special thanks to Professor Tal Arbel (McGill University; CIFAR AI Chair and core member of Mila – Montreal AI Institute), Professor Douglas Arnold (McGill University), and Dr Arman Eshaghi (University College London) for their insights.
References
[1] Messeri L & Crockett MJ Nature 2024; 627(8002): 49-58.
[2] Afzal HR, Luo S, Ramadan S, & Lechner-Scott J Mult. Scler. J. 2022; 28(6): 849-858.
[3] Wottschel V et al. NeuroImage Clin. 2019; 24: 102011.
[4] Nelson C A et al. JAMIA 2022; 29(3): 424-434.
[5] Pinto M F et al. Sci. Rep. 2020; 10(1): 21038.
[6] Falet J-P R et al. Nat. Commun. 2022; 13(1): 5645.
[7] Durso-Finley J et al. PMLR, 2022; 172: 1-20.
[8] Durso-Finley J et al. MICCAI, 2023; Cham: Springer Nature Switzerland.
[9] Rudin C Nat. Mach. Intell. 2019; 1(5): 206-215.
[10] Parikh RB, Teeple S, and Navathe AS. Jama 2019; 322(24): 2377-2378.