Spike sorting algorithm for intra-cortical recordings using unsupervised Bayesian decomposition

Nielsen, Martin Nøhr

4. term

Biomedical Engineering and Informatics, Master

2009

This thesis tackles the challenge of separating neural spikes from extracellular intra-cortical recordings without manual intervention. It adapts and evaluates a fully automatic, unsupervised spike sorting algorithm based on Unsupervised Bayesian Decomposition (UBD), originally developed for intramuscular EMG signals. The approach uses a Bayesian statistical model with a maximum a posteriori (MAP) estimator, is validated on both simulated and human intra-cortical recordings, and is compared with the classical unsupervised method Wave_Clus. UBD achieved performance similar to Wave_Clus, with average performance of 80.9% on simulated data and 83.2% on human data. In some simulated cases with short refractory periods, performance dropped to approximately 39%, whereas UBD showed strong capability in detecting and classifying overlapping spikes compared to Wave_Clus. Suggested improvements include retuning the TABU algorithm to handle more than three spikes per segment and extending the method to multiple channels to exploit inter-channel information.

Denne afhandling adresserer udfordringen ved at adskille neurale spikes fra ekstracellulære intracortikale optagelser uden manuel indgriben. Målet er at tilpasse og evaluere en fuldautomatisk, usuperviseret spike-sorting algoritme baseret på Unsupervised Bayesian Decomposition (UBD), der oprindeligt er udviklet til intramuskulære EMG-signaler. Metoden anvender en bayesiansk statistisk model med en maksimum a posteriori (MAP) estimator, valideres på både simulerede og humane intracortikale optagelser og sammenlignes med den klassiske usuperviserede metode Wave_Clus. UBD opnår næsten samme ydeevne som Wave_Clus med gennemsnitlig performance på 80,9 % (simulerede signaler) og 83,2 % (humane signaler). I visse simulerede tilfælde med kort refraktærperiode ses et markant fald i performance til omkring 39 %, mens UBD omvendt viser høj ydeevne til at detektere og klassificere overlappende spikes sammenlignet med Wave_Clus. For at forbedre metoden peges der på retuning af TABU-algoritmen til at håndtere mere end tre spikes pr. segment samt en multikanalsudvidelse, der kan udnytte inferens på tværs af kanaler.

[This apstract has been generated with the help of AI directly from the project full text]

spike sorting ; bayesian decomposition ; intra-cortical recordings

Download PDF
View record in AAU Student Projects