Introduction: Paediatric epilepsy represents a major clinical burden, particularly among children with developmental disorders. Surgical interventions have demonstrated benefit in reducing seizure burden. Identifying patients at risk of developing epilepsy and those likely to fail medical treatment could aid surveillance and earlier identification of surgical candidates. Artificial intelligence based methodologies offer a promising means of integrating high-level disease models to predict treatment outcomes.
Methods: We collected EEG data from a cohort (n = 42) of patients at high risk of epilepsy (Rett Syndrome). Patients were divided based on whether they developed epilepsy (no epilepsy: n = 18, epilepsy: n = 24), with the epilepsy group subdivided based on treatment response (treatment responsive: n = 16, treatment resistant: n = 8).
Electrophysiological features were characterised by deriving measures of spectral power distribution and functional network connectivity. These parameters were used to develop a series of classifiers. Performance was assessed using 10-fold cross-validation to investigate whether these models predicted epilepsy status and treatment response in new patients.
Results: A support vector machine trained on spectral power distribution identified asymmetry of power distribution as a marker of epilepsy, predicting epilepsy status with 69% accuracy (PPV 66.7%, NPV 72.7%).
A neural network classifier trained to integrate high-dimensional models of cortical network function based on inter-electrode coherence measures predicted treatment resistance with 87.5% accuracy (PPV 77.8%, NPV 93.3%).
Conclusions: Artificial intelligence methods offer a promising means of integrating many features into a predictive model in order to better predict epilepsy status and treatment outcomes. This facilitates the use of richer characterisations of disease in clinical decision-making, allowing treatment options to be informed by high-level models of the underlying disease. These approaches provide a potentially valuable method of risk-stratifying patients and of predicting those likely to fail medical treatment and require surgical intervention for epilepsy control.
Patient Care: Artificial intelligence approaches offer a means of identifying patients likely to develop epilepsy and to fail medical treatment based on routine non-invasive EEG recordings, allowing more targeted surveillance of at-risk patients and potentially allowing earlier definitive intervention
Learning Objectives: By the conclusion of this session, participants should be able to:
1) Describe the potential roles for artificial intelligence in integrating high-dimensional datasets to predict clinical outcomes
2) Discuss the characterisation of disease features as a means of predicting treatment response
3) Identify the applications of these techniques to guiding treatment surveillance and allowing personalised approaches to care