# Epilepsy research: Modulations in seizure states

## Overview

In another post, I described my work characterising seizures as pathways and investigating how these pathways change over time. One disadvantage of this approach is that it’s tricky to quantify what parts of pathways change from seizure to seizure. For example, we might want to ask whether seizures that occur at a certain time of day all share a certain feature. I therefore developed a complementary approach: describing seizures as state progressions, where each state in a seizure captures a particular pattern of brain activity. I could then quantify which states appeared in each seizure (state occurrence) and how long each state lasted (state duration).

I applied this approach to a unique dataset of chronic (spanning multiple months) recordings of brain activity in ten people with epilepsy, allowing me to investigate how seizure states change over days, weeks, and months. For example, the occurrence of many states changed over the course of a subject’s recording:

I was particularly interested in cyclical changes in seizure states, as seizure risk is known to vary over daily and multi-day cycles. I therefore extracted cycles in a biomarker of pathological brain activity and compared seizure states to these cycles. This analysis revealed that seizure states often vary cyclically.

Understanding how seizures change over different timescales could lead to new treatments that adapt over time to control different types of seizures.

## My contributions

This project was part of my PhD thesis; I was responsible for shaping the project’s direction, undertaking the analyses, and communicating the findings. The brain recording data was obtained and organised by collaborators from another research institution, who also provided feedback throughout the project.

## Data science approaches

**Signal processing:**I used signal processing techniques such as filtering to preprocess brain signals.**Network analysis:**I described seizure dynamics as the time-varying network interactions between pairs of brain regions.**Clustering and dimensionality reduction:**I used**non-negative matrix factorisation**as a soft-clustering method to extract recurring seizure states from time-varying seizure network interactions.**Time series decomposition:**I extracted cycles in brain activity using**empirical mode decomposition**, which decomposes a time series into oscillations at different frequencies. Unlike many other frequency decomposition approaches, empirical mode decomposition does not require the extracted cycles to be regular; it can capture prominent cycles even if the cycle period varies (e.g., if the cycle peaks every five to seven days instead of exactly every six days).**Circular and non-parametric statistics:**To characterise cycles in seizure states, I used a range of statistics, including**Wilcoxon rank sum tests**,**phase locking values**, and**circular-linear correlation**.