Epilepsy prediction and EEG biomarker detection
With the advent of affordable wireless headsets, EEG monitoring devices are set to become the wearable for the brain. Monitoring brain health for mental wellness will become as ubiquitous as activity tracking for physical fitness. NeuroPro’s advanced SaaS tools for mobile streaming of EEG data into the cloud enable the development of a big data community, to build collective knowledge to better understand the brain and neurological disorders.
Analysing such potentially vast data sets necessitates advanced algorithms designed to run intelligently in a standardised and unsupervised manner to deliver accurate and repeatable results. WINAM has been developed to address these future challenges in EEG research and has achieved promising results for epilepsy seizure detection and prediction.
WINAM is a patented n-gram based pattern detection algorithm. In general, the algorithm scans the input time-series (in this case EEG) by dividing it into overlapping segments of a user-defined length. These segments are further divided into a number of overlapping sub-intervals of another shorter, user-defined length. The amplitudes of the time-series in each sub-segment are quantised based on their amplitude to create n-gram pattern for each sub-segment.
Patterns that repeat more than twice are considered significant and the samples contained within each significant pattern are deducted from the total number of samples within each sub-segment. The “anomalies ratio” is then computed for each sub-segment by dividing the number of remaining samples (ie: those not included in any pattern) by the total number of samples in the sub-segment. This ratio is computed for each sub-segment and the resulting vector for each segment used as a feature vector for machine learning.
In the case of epilepsy seizure prediction and detection, models are built to machine learn the optimum WINAM input parameters to maximise sensitivity and minimise false alarm rates on a patient-specific basis. Input data containing the known location of seizures is used to train the algorithm. The algorithm is tested on data that was not used for training. The procedure iterates over all seizure and non-seizure data sets, isolating them for testing one at a time. Finally, it averages the results of each case to obtain the overall accuracy of the model for a set of input parameters. This training and testing is performed on each patient independently in order to capture patient-specific signal variations.
Self adaptive dynamic operation
NeuroPro’s data ecosystem infrastructure is designed to handle the continuous optimisation of WINAM’s biomarker detection and prediction algorithm such that it learns and adapts dynamically as the system captures missed seizures and false alarms.
Partner with us
If you are interested in using WINAM in your own research or in adapting it for use in the management of other neurological conditions, please get in touch.
Research Fellow in Medical Devices, Imperial College, London