AI gets real: using routine clinical data and Artificial Intelligence to predict worsening of Multiple Sclerosis despite treatment (AIMS)

Multiple sclerosis (MS) can be a severely disabling disease. Widespread use of MRI and revisions to MS diagnosis have enabled earlier identification and considerable progress in developing therapies for MS. Access to the treatments in MS has been improved recently, but although most people with relapsing MS can benefit from them and outcomes are improving as a result, no single treatment is right for everybody. Some people with MS will relapse, and over the long term will gather psychical and cognitive disability. Despite progress in assessing response to treatments, individual prediction of MS outcomes over the long-term is still inaccurate; the need for information on individualised long-term prognosis for MS patients' forecasting is frequently unmet.

A wealth of clinical and MRI data from patients with MS are acquired every year in clinical practice, but only part of these data are used for clinical decision making. MRI is paramount in MS diagnosis and monitoring, but most often the only feature of clinical use are the MS lesions. However, the structure of MS brains visualised on imaging is likely related to several aspects of MS biology. We propose that a set of structural characteristics extracted from the MRI brain scans of people with MS are related to biological changes which are meaningful to MS, and may therefore act as predictive markers for outcome. Computational imaging approaches using artificial intelligence (AI) have achieved successes in automatically quantifying lesions. AI-based classification of the scan's features referred to as 'radiomics' can provide more detailed characterisation than is possible by the naked eye and can offer the means to extract more information from the whole-image MRI brain scans.

Radiomics-based biomarkers (indicators) have shown success in cancer treatment, but are still in early development in MS. In this study, we aim to use whole-image brain MRI scans processed with AI techniques and detect the clinical and MRI profiles that predict accumulation of MS-related disability or cognitive impairment. We will take advantage of our MS clinic which is one of the largest in England, and the Nottingham MS Society Register. We will draw on a unique environment of research experts in MS, clinical trials, MRI, computational imaging and predictive modelling who work collaboratively with patients and carers within the NIHR Nottingham Biomedical Research Centre. We will use individual data about patients' clinical condition, their demographics and their scans, and analyse it by the means of AI. Using clinical information which patients have consented for us to use, and the MRI images before starting treatment, we'll train a computer to use mathematical models to predict whether a person's MS will determine accumulating disability or cognitive impairment over the long term. Furthermore, we seek to see if the profile can predict development of disability in other patient groups, by validating the models in large sets of MRI scans obtained from other groups of people with MS using different scanners. We will use a large group of patients from the United States, and also align with a clinical trial ongoing in UK and US, which compares treatments for MS with different strengths.

At least a third of people with MS starting on a first-line MS treatment require subsequent escalation to a stronger therapy. By identifying early, at diagnosis, who is likely to fare worse over the long term, we could offer them a more tailored treatment approach. This is a crucial step towards "personalised medicine", which means we'll be able to prescribe the right medication for the right person at the right time.

Multiple Sclerosis (MS) is a common cause of disability in the developed world. Despite progress in MS treatments, significant variability in disability accumulation remains. Prediction of long term outcomes and best use of all data acquired in clinical practice to define the profile of patients at risk for disability and cognitive impairment accrual, are both unmet needs. This project draws on the neurobiological heterogeneity of people with MS and aims to use computational models to identify specific profiles of people with MS who fare worse over the long term despite first-line treatment. We propose a time to event analysis for predicting progression of MS by comparing three different models and incorporating AI state-of-the-art techniques. We will exploit available real world data using AI and build on the experience of research groups at NIHR Nottingham BRC and the Nottingham AI program. We will use deep learning techniques (convolutional neural networks) to extract whole-image features from baseline MRI scans of more than 800 patients with MS, members of the Nottingham MS Society Register. We will use a classical machine learning based regression model, a deep learning model and compare their predictive accuracy with a classical statistical frequentist model incorporating clinical and classical MRI features. After internal validation, we will then externally validate the models in two large independent data sets of which one is a RCT which examines response to moderate versus high potency treatments in MS (DELIVER-MS). The results will help apply for further funding for large deep phenotype collecting (genetics,microbiome) biomarker studies targeted to these biotypes, and explore how they can be integrated into routine clinical practice.

The key driver behind this study is the patients that we see in our MS clinic who strive for early prognostic markers to inform their future, before irreversible disability occurs. Many of them consistently ask what information their clinical MRI scan provides for their prognosis; they show widespread interest in tools that could deliver individualised prognosis estimates and state that they would be willing to take part in studies to develop such tools. A recent UK nationwide in people with MS members of the MS Society Register study shows that most of people with MS wish to receive early information on long-term prognosis, but this need is unmet. Our PPI feedback confirms the "need to understand the progression" in MS and that the research addressing it is "vitally important" research. It would be of enormous benefit to patients if we could identify early imaging biomarkers to indicate prognosis at an early stage: this would impact how we treat and manage MS, and it could inform patients and carers to plan their future.
The impact on patient care of an easy-to-apply score that can be used at diagnosis to predict relevant disability and cognition outcomes, could be huge. Patients at risk for worsening disability could be offered more potent treatments at the outset and during a window of 'biological opportunity', and people with MS that have good prognosis with moderate efficacy drugs will be spared from exposure to high-risk treatments. It is for now impossible to say how much of the long term disability in MS can really be prevented with current disease modifying treatments, and how much is down to the natural history and individual biology. It's only by identifying these patients who are at risk of faring worse over time, that the impact of a treatment on MS course in a given individual can be clarified, and the risk-benefit trade-off of a treatment can be weighted.
The study could provide a baseline marker to inform clinical longitudinal studies of impact of specific types of MS treatments on cognitive impairment in MS. The whole-image and clinical modelling as predictor for cognitive impairment can be used in cognitive rehabilitation trials in MS, including studies of effects of early high-efficacy treatments on individuals at risk of developing cognitive impairment in MS.
The results of the study could be explored for implementation in the NHS. This would have impact on patient stratification for treatment and rehabilitation. Using routine MRI data for prognosis enables access to a continuously growing wealth of data, without any increase in the costs that the NHS already invests for MS diagnosis and monitoring. Implementing in NHS practice, a deep-learning based tool to extract whole-image features from routine scans to inform disease prognosis, can impact and improve NHS pathways of care in MS and has potential of exploratory application in other neurological conditions. NHSX could offer the direction, the setting and the potential of exploring the implementation of such an AI-based tool in NHS care pathways.
The model developed through refinement and external validation in two large independent data sets, of which one is a RCT, will proffer a large-scale proposal with a health economic investigation. The radiomic profiles obtained will be fed into the Nottingham BRC imaging network to enhance impact and will be further explored through collaborations with the MS-PATHS and MAGNIMS links for the wider community of researchers, scientists and trialists.
The validated clinical-radiomic models would be used to define populations at risk. These populations could be further subject to extensive patient profiling, including genomics, pharmacogenomics and fully characterised biomarker profiling. The results of this study could be fed into very-large data prospective studies currently ongoing in Europe and USA of biomarker profiling and precision medicine in MS.