Our work in Brain monitoring involves with application of learning based analytic techniques to assist clinicians on decision making leading to efficient treatment schemes. This involves development and effective usage of machine learning/statistical techniques and derivation of knowledge representation models based on expert domain knowledge. Our strong collaboration with colleagues at the Mayo Clinic provides us access to clinically collected data from patients with variety of neurological diseases. In the following, we provide two examples of our work using timeseries iEEG data collected from patients with epilepsy.
Temporal prediction of epileptic seizures
Epileptic seizures are traumatic events which cause a constant anxiety in epilepsy patients about when the next seizure is going to strike. Because of this reason, epilepsy patients generally take continuous medication and refrain from standard activities such as driving, swimming, etc. The ability to predict the occurrence of a seizure ahead of time will enable, timely medication and treatments preventing seizures such as electric stimulation. Hence, we developed a seizure prediction framework using real iEEG data collected from canine subjects who had epilepsy. This framework, apart from the ability to predict epileptic seizures on average 90 minutes ahead of time, could also identify patient specific pre-seizure signatures which enabled individualized treatment of the condition.
Spatial localization of seizure generating brain region
Surgical removal of seizure-generating brain tissue can cure epilepsy in patients who do not respond to medications. However, identifying seizure-generating regions is difficult and fails in many cases. We developed a fully unsupervised and automated approach to seizure focus localization using a Bayesian filter. This method uses a spectral domain feature, Power in Bands (PIB). PIB is extracted from inter-ictal (non-seizure) intracranial EEG recordings of patients with focal epilepsy to differentiate normal and abnormal brain regions. Experiments show that using a Bayesian filter for capturing temporal properties of the iEEGs recorded from epileptic brains remarkably improves localization accuracy. Our study also reaffirms that high-frequency oscillations and inter-ictal spikes are useful inter-ictal biomarkers of the epileptic brain, and PIB, which could be implemented with relatively low computational burden, performs as well as the standard bio-markers when used in this setting.