N-CODE Stakeholder-Led Study Group April 2025

Challenge presenter: Andy Wills (University of Plymouth) 

The economic costs of autism in the UK exceed £32 billion per year. Quick, accurate and cost-effective diagnosis can help, but ASD diagnosis is currently none of those things. NHS waiting times are long. Private diagnoses cost in the order of £3000. The process of diagnosis is time consuming, labour-intensive, and requires the expertise of a very limited group of highly qualified individuals. There are also emerging concerns of the quality of diagnosis and diagnostic overshadowing. This is because ASD is diagnosed by three cardinal symptoms whose aetiology can be diverse, expertise for which is not easily available. Identification of biomarker to robustly screen or diagnose ASD would be a “gamechanger”, for patients, clinicians, services and for the health and social care economy. 

The University of Plymouth has developed proof-of-concept research which demonstrates that AI-supported analysis of resting-state functional MRI (rs-fMRI) brain scans can be used to diagnose autism with 98% accuracy [Ref] and makes use of large, publicly available datasets.  

Questions Tackled During the Meeting: 

The overarching challenge was how to go from proof of concept to a workable, cost-effective, community-based commercial product? 

To address this, these challenges were considered:

• the publicly available data, which are essential for training the AI, are licensed for research only and the proof-of-concept system is a binary classifier, i.e. it can only respond ‘autistic’ or ‘neurotypical’ (may increase false-positives). Thus, what is and is not permissible use of research-licensed data, and what are the practicalities of collecting large-sample data sets, in the order of 1000s of scans based on prior research. 
• Issues relating to the fast, cost-effective, and accurate application of the proof-of-concept to real-world clinical settings. The proof-of-concept AI uses resting-state functional MRI (rs-fMRI) as its input. This information is not collected in clinical settings, where almost all uses of MRI are structural scans. Can the proof-of-concept approach be developed such that it is robust, quick, and turn-key reliable in a clinical setting, or if the findings can be translated to use structural data which are readily available. 
• Improving community-based diagnosis: availability and portability of scanning technology.  The first step towards broader accessibility is the introduction of NHS Community Diagnostic Centres; an important development that nonetheless brings its own challenges in terms of translating work so far done in high-field strength scanners with research-active radiographers to the low field-strength, less research intensive, multi-manufacturer context of community diagnosis. But, for wide community-based use, we need to go further. How do we do this? One option would be to use portable MRI, or explore using FNIRS rather than fMRI. The technology already exists and is already about as portable as EEG.

You can view the full challenge submission here. 

Challenge presenter: Dave Cook (E-Mind Ltd)  

Neurorehabilitation is a health strategy that aims to enable people experiencing or likely to experience neurological disorders to achieve and maintain optimal functioning including in mobility, vision, hearing and cognition. Neurological issues, including stroke and dementia are significantly increasing globally, supported by limited clinical resources, requiring solutions that address these problems in a highly scalable, cost effective manner. The use of digital technology and large language models offer significant opportunities to provide a more inclusive service, maximising the impact of remote neuro-management and community care services. 

Problem 1: Addressing Equity, Diversity and Inclusion (EDI) in Neuro-Diagnosis and Treatment

Where large language models offer highly automated solutions to personalise support and rehabilitation services, enabling patient care outcomes to be optimised. The three groups who frequently face limited access to neurorehabilitation services are: (i) individuals living in remote areas, where neurorehabilitation for rural populations is particularly challenging, (ii) ethnic and linguistic minorities where diagnosis of cognitive abilities is often inaccurate and (iii) individuals with disabilities and varying speech and/or motor abilities. 

Problem 2: Poor Patient Management & Fragmented Integration Between Health and Social care

For many neurological conditions, self-management is the key element to recovery and improving patients’ autonomy. How could effective usage of data visualisation be used to improve communication between practitioners and patients. The use of machine learning and mathematical modelling techniques could provide rapid auto-triage of patients, predict patients’ recovery outcomes for care plans, personal AI assistants and personalised scaffolding for rehabilitation, and resulting in more effective care plans.  

Problem 3: National Shortage and Inefficiency of Neurorehabilitation Services

How can technology help addressing these growing issues in neurorehabilitation to alleviate burden on the healthcare system?  

Questions Tackled During the Meeting: 

• How can technology such as ML/AI assist clinicians in providing remote, community care services? 

• How can AI support “information transformation” or “content matching” in order to present information flexibly to support patients’ specific needs? 

• Can other technologies (e.g. logic programming) create synergistic effects to LLM performance, reducing known LLM risks such as hallucinations? 

• What are the route to market options for this type of product? 

• How can the UK health ecosystem support the testing and deployment of remote rehabilitation services? 

• What is a sensible roadmap for the different stages of product development to maximise impact on a national scale? 

You can view the full challenge submission here.

Challenge presenters: Brandan McLean (Royal Cornwall Hospital NHS Trust) and Lance Vincent Watkins (Swansea Bay University Health Board)  

Sudden Unexpected Death in Epilepsy (SUDEP) is the leading cause of death for people with epilepsy but remains poorly understood. The main risk factor of SUDEP is poorly controlled generalised seizures. But beyond such clinical risk factor, there are no established quantitative biomarkers of SUDEP. With no methods existing to effectively predict or prevent SUDEP risk, predicting or preventing SUDEP remains challenging, highlighting the need to understand its aetiology and identify risk factors and biomarkers. This proposal brought together 2 related challenges.  

Developing an Artificial Intelligence Framework for Identifying People at Risk of SUDEP (Watkins)  

SUDEP often occurs following a generalised tonic-clonic motor seizure (GTCS), more frequently during sleep. GTCS has a strong link with distinctive changes in ECG and EEG signals, suggesting the interaction between seizures and autonomic nervous system function. Cardiac changes have been observed before death in SUDEP caused by GTCS events. Continuous ECG monitoring for people with frequent GTCS can identify high-risk individuals. GTCS are also associated with specific EEG patterns reflecting aberrant brain activity during seizures. Consequently, ECG and EEG, are essential tools for determining the risk of SUDEP. This challenge aims to develop an AI model based on identified SUDEP risk markers that can be applied to EEG and ECG recordings to help predict SUDEP risk. 

Questions Tackled During the Meeting:

• Does the analysis of the GTCS frequencies help to identify the risk factors for SUDEP?  

• Could changes in EEG and ECG signals help with understanding of the aetiology of SUDEP and identifying its risk factors?  

• Is it possible to develop an AI model that helps reduce the fatality rate of SUDEP by identifying its risk factors? 

• What are the innovative AI techniques that can be investigated to accurately identify individuals at risk of SUDEP based on analysing changes in EEG and ECG signals?  

• What is the required level of performance for the developed AI-model to be clinically feasible to identify the risk factors for SUDEP?  

Exploring SUDEP Biomarkers from Electrophysiological Data (McLean)

Previous studies suggested that interictal EEG recordings are altered in SUDEP. Brain networks from EEG data have shown to be associated with seizure propensity, presenting powerful tools to support epilepsy diagnosis and management. Understanding the role of brain networks in SUDEP is essential to comprehend its mechanisms. Despite some studies indicating that SUDEP is associated to altered connectivity, little is known about the relationship between brain networks and SUDEP. 

Networks from peri-ictal EEG and resting-state fMRI imaging were shown to have different properties in SUDEP. Despite the practical limitations of the biomarkers presented in those studies, their results indicate that interictal EEG networks can provide viable biomarkers of SUDEP. Network-based biomarkers of SUDEP from non-invasive data would allow clinicians to develop optimized personalised care plans to reduce mortality and improve life quality. It would also advance our understanding of the underpinning mechanisms of SUDEP in risk groups (e.g. Dravet syndrome), ultimately supporting the development of preventive intervention.  

Currently risk of SUDEP is assessed using standardised questionnaires based on clinical risk factors derived from retrospective studies, and do not give more than a general picture with no weighting attached to the individual risks identified. More specific biomarkers have not been explored in the at-risk groups. There is a need for a more objective biomarker to aid in identifying at risk individuals. 

EEG data on individuals who have died with SUDEP is available through the Royal Cornwall Hospital NHS Trust and Cornwall Coroner. These data are available for analysis using AI techniques, with matched controls, to explore published biomarkers. The ECG single channel data from the same EEG recordings will also be analysed, together with any other ECG data that is available in a digital form.

Questions Tackled During the Meeting:

• What is the relationship between brain networks and SUDEP? 

• Can background EEG and ECG be used to evaluate the risk of SUDEP susceptibility? 

• Can biomarkers provide serial information on changing risks? 

• Are biomarkers superior to, or complementary to clinical risk factors already identified? 

You can view the full challenge submission he

Challenge presenter: Omair Ahmed (Neurometry Ltd)

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by impairments in social interaction, communication, and repetitive or restrictive behaviours. Despite advances in understanding the neurobiological and cognitive mechanisms underlying ASD, its diagnosis, particularly in adults and across diverse populations, remains challenging. Traditional assessment methods such as questionnaires or resource intensive approaches like EEG and fMRI are limited in accessibility and scalability. 

Emerging evidence suggests that interactive, touch-based games can provide valuable insights into cognitive and motor domains, offering a promising alternative for ASD assessment. For example, motor performance variability in tablet-based games has been linked to early indicators of ASD. Such gamified approaches, designed to address specific cognitive domains like attention, memory, and motor execution, could offer cost-effective, engaging, and culturally adaptable tools for early screening and long-term monitoring of ASD. Neurometry’s existing digital health platform and preliminary tablet app architecture will be available for data collection. 

This challenge seeks to explore how a gamified, multi-domain assessment can be developed and validated as an effective tool for ASD research and intervention.

Questions Tackled During the Meeting: 

• How can we design and validate a gamified tool that addresses key cognitive and motor domains (e.g., attention, problem-solving, motor execution) for ASD assessment? 

• What methods can ensure cross-cultural adaptability and accessibility of this tool? 

• How can game-based data (e.g., reaction times, error rates, variability) be used to develop robust digital biomarkers for ASD diagnosis? 

• Can computational neuroscience models effectively integrate behavioural and motor data to predict ASD-related traits or severity? 

• What ethical considerations should be addressed for deploying game-based assessments in neurodevelopmental research?  

• What regulatory pathways are suitable for certifying such a tool as a clinically valid Class 1 device or Health Performance tool? 

You can view the full challenge submission here.