Full Program »

Big Data, Artificial Intelligence and Machine Learning In The Diagnosis and Management of Epilepsy

Administrator: Richard Chin

Machine Learning to improve genotype-phenotype matching and identification of cognitive/behavioural problems
Richard Chin

Two of the major challenges of epilepsy management are (1) genotype-phenotype matching and (2) early identification of cognitive/behavioural problems.  Variance of unknown significance is a challenge often faced by clinicians when a genetic abnormality is found, but the clinical significance is uncertain. Current databases such as OMIM require manual curation and only use 10% of available published data which limit scalability. A novel machine learning automated approach using Cadmus, a bespoke literature retrieval system, will be presented demonstrating an improvement compared to OMIM in genotype-phenotype matching for single gene disorders, many of which have epilepsy as a feature. Identification of cognitive/behavioural problems usually requires labour intensive assessments by psychologists and are not universally accessible. In comparison, EEGs are often used in epilepsy diagnosis worldwide. Innovative methods of identification of such problems using data from routine EEGs will be shared and discussed. MRIs are also often used in epilepsy diagnosis and we will provide data on network generation using standard sequences to predict cognitive problems. Finally, developments in identification of cognitive problems using multimodal modelling from EEG and MRI data will be discussed.

 

Multicentre Epilepsy Lesion Detection project: bringing AI into epilepsy presurgical planning 
Konrad Wagstyl

Focal cortical dysplasias (FCDs) are a common cause of drug-resistant epilepsy that can be challenging to identify on MRI. However, when detected, they can often be treated effectively through surgical resection. The Multi-centre Epilepsy Lesion Detection (MELD) Project is an international collaboration dedicated to improving the detection of FCDs in patients with drug-resistant epilepsy. In this talk, I will present our work on developing and training a neural network to detect FCDs using a large cohort of 1015 MRI scans from 22 epilepsy centers worldwide. Our algorithm achieved an overall sensitivity of 67% and was able to detect 63% of lesions previously considered MRI-negative. I will also discuss how our pipeline generates individual patient reports that identify the locations of predicted lesions and their imaging features and relative saliencies to the classifier. Finally, I will share updates on how the MELD pipeline is being integrated into presurgical planning for epilepsy at Great Ormond Street Hospital (GOSH) and other centers around the world.

 

Generative models of brain dynamics in epilepsy
Richard Rosch

Epilepsy is characterized by abnormal brain dynamics, which can be measured with increasing sophistication in patients and preclinical models. Concurrently, the growing availability of detailed genomic and neuroimaging diagnostics aids our understanding of the synaptic and network constraints under which these pathological brain dynamics emerge. Generative models of brain dynamics facilitate the investigation of how changes in synaptic efficacy or network topology may influence abnormal brain activity. While these models often identify qualitative features of the dynamic landscape, they can be challenging to relate to specific empirical data, particularly at the single-patient level.

The development of efficient artificial intelligence (AI) methods for fitting generative models to empirical data now enables the connection between quantitative observations from preclinical models and patient data to models of epilepsy pathology. Here, we will explore how AI enables the inference of constraints underlying abnormal dynamics in brain recordings from animal models and patients. We will illustrate this using specific examples, ranging from whole-brain single-cell calcium imaging in zebrafish to large-scale normative datasets of intracranial EEG data. We will focus on certain Bayesian model fitting techniques, such as dynamic causal modelling, and demonstrate how these AI methods are poised to impact clinical practice in the near future

 

Responsible AI in paediatric neurology; smartphone video and clinical metadata facilitating rapid diagnosis & treatment
Sameer Zuberi

Deep learning models for detection of epileptic seizures spatiotemporally from video may develop from the interface of fundamental computer vision, human motion analysis, machine learning research and interdisciplinary working between clinicians and data scientists. Early research has primarily used highly controlled video in epilepsy monitoring units or using multiple cameras. Research focusing on analysis of uncontrolled carer recorded smartphone video will be challenging but has the prospect of leveraging data from 6.9 billion smartphones and more accurately reflecting community needs/clinical resources worldwide. Integrating a research programme through a national ethical framework has allowed our team to develop a neurology video research database from a clinical platform vCreate Neuro (www.vcreate.tv/neuro). The application has > 27k videos and is growing by > 800/month. Established in >70 services, projects are being established in high and restricted resource settings in Europe, N America, Asia, Australasia, the Pacific, and Africa.  Videos are uploaded to a secure web-based platform with associated metadata and classified by clinicians, facilitating a curated clinical and research database. We will discuss the concepts of responsible AI as applied to video of children, research methodologies including annotation, 2D & 3D skeletal pose sequences and how an AI algorithm can be integrated into a diagnostic and management pathway and adapted to local models of healthcare