Hasso-Plattner-InstitutSDG am HPI
Hasso-Plattner-InstitutDSG am HPI

Self-prediction of epileptic seizures by affective computing using Electroencephalogram (EEG) sensor

Sidratul Moontaha

Chair for Digital Health - Connected Healthcare
Hasso Plattner Institute

Office: Campus III Building G2, Room: G-2.1.21
Email: sidratul.moontaha(at)hpi.de
Links: Homepage  

Supervisor: Prof. Dr. Bert Arnrich


My current research work focuses on affective computing using EEG sensors. The other sensor modalities include i.e., Photoplethysmogram (PPG) sensor.  The subprojects expands from classifying cognitive load to predicting epileptic seizure. In the following sections, my current project will be followed by the past projects.  


Affective state classification during relaxation and cognitive performance


Human brain integrates emotional experience with cognition, handles the emotional reaction in response to the external stimuli, and also works as a biological basis of emotions that
stores fear and anxiety. Therefore, analyzing brain signals for cognitive load classification is more straightaway since cognition takes place in the human brain. Moreover, nowadays, the EEG recordings are becoming more convenient to use in daily life through fewer electrodes, unlike the in-hospital setting EEG recordings used before. These latest developments in non-invasive and portable consumer-grade wearable Electroencephalography (EEG) have earned special interest from the research community. 

On the other hand, for an objective measure of cognitive load, the datadriven physiological measures are mostly skin conductance, electrooculography, anterior cingulate cortex signal, blood volume pulse, temperature, audio data and magnetoencephalography. The increasing popularity of wearable devices measuring galvanic skin response, eye activity, respiration, and heart rate has become increasingly prominent for less obtrusive online measurement of cognitive load. However, these measures still suffer from limitations from both confounding factors and noise. Therefore, in this project, we use the non-obtrusive EEG sensor to measure the cognitive loads. 



Study Design

In order to conduct the experiment, two laboratory set ups has been designed; one with coginitive load session and another with relaxation task session. These two randomized controlled session will provide us the effects of the tasks on the physiological parameters. We recruited 11 participants to participate in two laboratory studies of arroximately 2 hours each session. In each session we provided them two hand worn wearable sensors consisting of PPG, EDA and IMU sensosrs and one headband sensor consisting of EEG sensors. In addition to that we record the audio signals of their certain moods. After the sensor calibration in both the sessions, the participants were shown a relaxation video to get over any previous short term stress, for example, any activities that might lead to a higher heart rate. Participants also fill out a pre and post mood questionnaires named Positive and Negative Affect Schedule (PANAS) and rate their mood in terms of valence and arousal in affective sliders. In the particular coginitive task session, the participants performed n-back task, stroop test task, and reading span task. Between each task they rated the difficulty of their task with Nasa task load questionnaire (NASA-tlx).

Data Analysis: Pilot Study

The above mentioned data is in a process of labelling according to the questionnaire. While collecting this data another small pilot study has been conducted with a portion of the above mentioned study design where only the brain activity of 9 participants were recorded with EEG electrodes while performing three difficulty level of cognitive tasks followed by a recovery phase as depicted in figure below. The EEG signal needs significant ammount of preprocessing to remove the physiological artifacts e.g., eye blinks, eye movements, head movements, heartbeats and non-physiological artifacts e.g., power line interference, electrode artifacts due to poor placement of electrode. Therefore, a low pass Butter-worth filter
of the sixth order at 44 Hz is applied to remove higher frequency noise caused by the muscle activation following a notch filter at 50 Hz to remove the power line interference which corresponds to the utility frequency in Europe. Besides, a high pass filter is applied at 0.5 Hz for removing the low frequency physiological artifacts. While performing filtering it was ensured that the valuable information from all frequency band of the EEG signal is preserved. An additional filter at 20 Hz is applied to remove the movement artifact. The normalization of the data is performed from the baseline of the eye closing session for each subject. The labelling of the data was obtained from the Nasa-tlx questionnaire filled up at the end of each task. Two classifciation task was performed based on the 0 to 100 point scale based on the label. Firstly, binary classification between high load and low load. Secondly, tertiary classification between low, medium, and high cognitive load. Instead of the conventional hand crafted feature extraction methods, we used automated feature extraction by Echo-State Network (ESN) and compared the performance to Sparse Auto Encoders (SAE). The result shows that with the same amount of preprocessing applied
for both the methods, the ESN achieves 36% higher accuracy in binary classification and 28% higher in tertiary classification. Additionally, ESN can increase the classification accuracy by 7% for binary classification and 23% for tertiary classification without significant preprocessing of the data. At the same time, the sparse auto-encoder needs more preprocessing of the signals. The results are submitted to the Thirty-Fourth Annual Conference on Innovative Applications of Artificial Intelligence (IAAI-22).

Analyzing the effects of anti-epileptic drugs on seizure frequency

This project aims at developing and implementing a flexible methodology for modelling arbitrary count time series by employing a state space modelling approach and at designing appropriate control strategies based on the modelling results. The focus of the intended applications lies on biomedical time series; in particular, the task of monitoring and controlling the therapy of patients suffering from epilepsy by AED medication will be addressed. This application will serve as a realistic testbed for the developed methodology.

Quantitative methods from system identification and control will be introduced into an important field of clinical therapy planning. We aim at making optimal use of the available data, i.e., of the response of the patient to previously administered medication. This project is based on the hypothesis that the quantitative approach to analysing this data based on time series analysis and state space modelling will provide superior results, as compared to the current standard given by the direct least-squares regression approach or by visual inspection and subjective decision based on the experience of the physician conducting the treatment. As a result, the physician will be provided with a tool for automatic monitoring of ongoing treatment and for suggesting changes and adaptations. The minimum data set size that is required for such tool to function reliably will be investigated, possibly by a suitable simulation approach. We expect that our approach will offer a solution also for cases that are difficult to assess, e.g.,  given simultaneous administration of several drugs with time-varying dosages.

In future work, also additional data besides the daily counts of seizures may be added to the analysis, such as data on side effects of the medication.

To achieve these goals this project aims at improving and extending the available methodology for nonlinear Kalman Filtering and for estimation of model parameters by maximisation of the likelihood given count time series of limited length, as are common in many biomedical applications. For this purpose, recently proposed algorithms, like square-root Kalman Filtering and Kalman Filtering based on Singular Value Decomposition, will have to be generalised for the nonlinear case. Also available algorithms for numerical optimisation will have to be adapted to the particular situation by employing recent developments, such as square-root versions of the Expectation Maximisation algorithm.