Reporte interno proyecto Neurosense: Análisis de oscilaciones de alta frecuencia (HFO) en registros neuronales

DAMIÁN DELLAVALE

2022-01-01/2022-07-31

Informática (software)

Tecnologia sanitaria y curativa

Reporte interno elaborado en el marco del postdoctorado realizado en el- Institut de Neurosciences des Systèmes(INS, UMR1106), Aix MarseilleUniversité, INSERM, Marseille, France.- APHM, Timone Hospital, Epileptology and cerebral rhythmology, INS, Inst Neurosci Syst, Marseille, 13500, France.- Université de Rennes 1, LTSI and INSERM, U1099, Rennes, F-35000, France.Resumen:Several waveform shapes observed in the brain activity like spikes, ripples and fast ripples,have been proposed as biomarkers of epileptogenic tissues. The clinical relevance of thesebiomarkers arises from their capacity to improve the delineation of the epileptogenic zonewith an impact in the outcome of surgical interventions. A more fundamental goal is to bettercharacterize the dynamics of epileptogenesis. In spite of the numerous proposed methods,detection and classification of epileptiform waveforms remain challenging and are openproblems in clinical and translational neuroscience. Our goal is to propose a method withminimal supervision that can detect and classify events of interest in electrophysiologicalrecordings, based on data driven basis functions.The proposed framework is constituted by a data driven feature extraction module integratedin a user guided adaptive filtering loop. The feature extraction is based on dimensionalityreduction methods in charge of producing time-frequency basis functions characterizing thewaveforms of the events and leading to scores used in a clustering stage. The filter kernelscan be initialized with standard band pass filter kernels or with the basis functions associatedto a pre-selected set of events. In the successive iterations of the filtering loop, the input datais filtered using the basis functions which effectively capture the features characterizing thewaveforms of the events present in the data. The supervision of the adaptive filteringprocess can be done on-line by choosing the desired features in each iteration of the filteringloop or, by defining a priori the strategy for updating the filter kernels. In each filtering loopiteration, the event detection is performed by thresholding the filtered data. The amplitudethreshold is automatically computed using the local false discovery rate technique [Efron2004]. Importantly, the resulting algorithm has a reduced number of free hyperparameters.The proposed strategy is weakly supervised in the sense that it does not requires big labeleddatasets for training. Instead, it allows informed user interventions to guide the adaptivefiltering process on the basis of the time-frequency features effectively observed in the data.Preliminary results using synthetic [Roehri et al. 2017] and experimental recordings inepileptic patients and a mice model of epilepsy, are presented as a proof of conceptdemonstrating the clinical relevance of our method.