Electroencephalography is a method to record the brain’s electrical activity and one of the most common tools used to study brain functions and neurological disorders. Multiple electrodes placed along the scalp measure voltage fluctuations resulting in a multi-channel signal measurement, the electroencephalogram (EEG).
For example, EEG examinations are performed in the diagnosis and monitoring of epilepsy and sleep disorders. Epilepsy monitoring is conducted to characterize epileptic seizures and detect interictal (between seizures) epileptiform activity. This activity consists of wave patterns such as spikes, sharp waves and slow waves that can appear individually or in groups.
Similar interictal epileptiform abnormality in combination with abnormal signal slowings may also appear in schizophrenic patients treated with antipsychotic medications such as clozapine.
The inspection and interpretation of clinical EEG is done by experienced medical professionals. The analysis is conducted based on the properties of the EEG signal such as amplitude, shape and frequency. Additionally, the conditions, the patient’s clinical history and medications are considered. The human-based visual inspection is still the gold standard in EEG interpretation. However, visual inspections are time-consuming since the specialists analyze every part of the recordings. A typical clinical recording may last anywhere from 20 minutes to several days, which produces a huge amount of data and makes the visual analysis inefficient.
Also, new devices such as mobile EEGs and brain-computer-interfaces will gather even larger amounts of data over longer periods of time. Therefore, automated analysis is desirable and existing algorithms should be further developed and tested.
A number of methods for automatic detection of relevant EEG events is implemented in the clinical software. The performance of these algorithms however is still far from satisfactory and no direct adjustment is possible because of the lack of open source code.
In order to be able better control the output of automatic event detection we conduct research on methods that help in the detection of abnormal signal events. Based on publicly available EEG databases we developed and evaluated machine learning and signal processing methods. Currently we explore the possibilities of deep learning methods (recurrent and convolutional networks) and working on combining text mining (patient’s data) with the EEG signal analysis in order to gain better insights.