Detecting Bursts

The simplest classification was a burst that occurred in a single hill of the EHV analysis

signal (Fig. 4.3A). The rising height, h1, of the hill was measured from the leading trough to the

peak. Similarly, the falling height, h2, of the hill was calculated from the peak to the lagging

trough. The width of the hill, w, was calculated from the leading trough to the lagging trough. If

both the rising and falling height-to-width ratios, !!

!, exceeded a user-defined value, called the

Rising Peak >= Tonic Thresh. Next Peak Yes No Hill-Valley STD >= Tonic STD Minimum Signal Value >= Tonic Min. Yes No Yes Complete TONIC ACTIVITY No

TONIC ACTIVITY LOOP

Signal Conditioning Burst Detection Loop Next Peak

A. Tonic Spiking Detection C. Tonic Activity Detection Algorithm

B. Detecting Bursts & Tonic Spiking

Figure 4.4. Extended Hill-Valley classification of tonic activity

Spike raster plots and instantaneous frequency graphs illustrate tonic spiking activity. The EHV analysis detects features in the EHV analysis signal. Burst events are indicated above each raster plot by thick bars while bouts of tonic activity are marked by lines with barbs on either end. Results of visual inspection, CMA, and PS algorithms are shown for comparison. (A) Onset of tonic activity is identified when the height-to-width ratio (h/w, red arrows) of a hill exceeds the user-defined “tonic threshold.” Termination of tonic activity is determined by

either a trough that falls below a minimum activity level, MIN, or by an excursion of a peak or

trough that exceeds the Hill-Valley STD, +STD. (B) The EHV algorithm is capable of

discriminating bursts and tonic spiking that are immediately adjacent. The instantaneous firing frequency is shown below the raster plot to assist in visual identification of bursts. (C) A series of criteria are used to identify tonic activity as highlighted by the thick outlines and arrows in a reduced flow diagram of the Extended Hill-Valley analysis algorithm. A full flow chart of the EHV algorithm can be found in Appendix A, Fig. A.1.

93

multiple hills (Fig. 4.3B), the height-to-width ratios were measured from the highest peak in the sequence to the leading and lagging troughs of the entire sequence. Because heights were measured from trough to peak rather than the absolute height of the peak, bursts were detected independent of different baseline firing rates.

Visual inspection of an in vitro spike train that was spontaneously bursting (Fig. 4.5A)

identified 25 bursts and no tonic activity. The EHV and CMA algorithms detected 15 and 14 bursts, respectively, while the PS algorithm detected 61 bursts. Full analysis results for sample data are shown in Fig. A.3 of Appendix A and short examples were selected in Fig. 4.5 to highlight the differences of algorithm performance. Analysis of simulated rhythmic bursting (Fig. 4.1A), identified 36 bursts with an average duration of 4.88 +/- 0.15 s by visual inspection. EHV detected 35 bursts that had a similar average duration of 4.62 +/- 0.12 s and CMA detected 44 bursts with an average duration of 3.81 +/- 0.27 s. PS detected 239 burst events with an average duration of 0.24 +/- 0.01 s.

EHV detected fewer bursts in the in vitro spike train because some events that were

visually identified had a lower spike frequency and the corresponding hills of the EHV analysis signal did not meet the burst threshold criteria.

The shape of the ISI distribution used by CMA analysis resulted in an unfavorable selection of ISI thresholds for simulated rhythmic bursting. CMA uses two ISI thresholds to identify the onset and termination of bursts. In the case of simulated rhythmic bursting (Fig. 4.2A, black and white diamonds), CMA detected spikes at the beginning or end of a burst that had an ISI greater than the threshold value but were still visually identified to be part of a burst.

The burst threshold ISI for in vitro bursting was 5.8 s and was sufficient to classify most of the

bursts in the spike train. A secondary threshold was used to add spikes to the beginning or end of a burst in case they had a lower ISI but were still part of an event. Because of the shape

of the ISI histogram the secondary threshold was 22.0 s and resulted in several of the bursts being joined into single events.

A larger number of bursts was detected by PS than by the other two algorithms. Because PS identifies bursts based on the likelihood that a sequence of spikes occurs by

chance, spike trains with low ISI variability resulted in poor performance. For spontaneous in

vitro bursting, the average firing frequency of spikes in visually identified bursts was 42.56 +/- 1.63 Hz while the average population firing frequency was 42.35 +/- 0.71 Hz. In the case of simulated rhythmic bursting the visually identified events had an average spike frequency of 55.07 +/- 1.07 Hz while the population average of the entire spike train was 54.53 +/- 0.39 Hz.