Fast Automated Interictal Spike Detection in iEEG/ECoG Recordings

Using Optimized Memory Access

Filip Kesner

1,2

, Jan Cimbalnik

, Irena Dolezalova

, Milan Brazdil

and Lukas Sekanina

Faculty of Information Technology, Brno University of Technology,

Brno, Czech Republic

International Clinical Research Center, Center of Biomedical Engineering,

St. Annes University Hospital, Brno, Czech Republic

1st Department of Neurology, St. Annes University Hospital and Medical Faculty,

Masaryk University, Brno, Czech Republic

1 MOTIVATION

Interictal spikes have been established as an impor-

tant biomarker in surface EEG and intracranial iEEG

recordings for some time (Staley et al., 2011). Spikes

are used for clinical practice and research of epilepsy,

ADHD and also in other areas (Barkmeier et al.,

2012a). Although the gold standard for interictal

spike detection has been and still mainly is a man-

ual evaluation, it has been shown that higher consis-

tency of results can be achieved by automated detec-

tion algorithm (Barkmeier et al., 2012b). Detection

algorithms can save enormous amount of work for re-

viewers and provide a faster data analysis for research

or even clinical practice.

2 OBJECTIVES

Computational efﬁciency is not so important when

recordings are processed from only a few channels

and a real-time detection is not necessary. Example of

those would be recordings from rodents (Ovchinnikov

et al., 2010). However, when processing intracranial

recordings from humans, in as much as 150 channels

with 5 kHz sampling rate, which are in average 30

minutes long, computational time requirements gain a

great deal of importance. While several terabytes (just

our institution) of such recordings are available for

processing, a detection algorithm has to be designed

to allow fast ofﬂine processing of intracranial record-

ings or even a real-time detection over at least hun-

dreds of channels simultaneously. In order to process

large signal data, the memory access is often a cru-

cial bottleneck for CPU processing, which puts high

requirements on effective cache utilization, to reduce

the access frequency to a slow main memory. The

goal of this paper is to propose an efﬁcient spike de-

tection algorithm, particularly, the ﬁrst level detector.

3 METHODS

3.1 Data Acquisition

Signal data which have been used for evaluation of

this detection algorithm were recorded from patients

suffering from pharmaco-resistant form of epilepsy.

The areas of brain where stereo electrodes have been

positioned vary through patients. This variability of

signal source is useful for algorithm testing, providing

a complex good-quality dataset. Signals have been

recorded approximately for 30 minutes each in 129 -

150 channels. Recordings also contain 6 non-iEEG

channels such as ECG, EOG, and calibration signals,

which can be omitted from processing. The recording

device records the data with 25 kHz sampling rate,

subsequently down-sampling them into 5 kHz range,

which is still relatively high, but it is necessary for

detection of other possible biomarkers, such as HFOs.

To illustrate the enormous size of such data

recordings, the channel size is expressed as:

channel size = 5000Hz*(30min * 60sec)* 4bytes

where the average recording ﬁle contains 150 such

channels, resulting into the ﬁle size of 5.4 GB, which

can be estimated by the following formula:

file size = 36MB * 150channels

Recordings of intracranial EEG are huge ﬁles and ter-

abytes of such data are available for processing (just

at our institution), which should be done by the pro-

posed algorithm for one 5 GB ﬁle in tens of seconds

instead of tens of minutes, as it has been done before.

3.2 Detection Algorithm

The detection algorithm has been designed to be mod-

ular, thus allowing the choice of how many modules

will be employed in detection. This approach en-

ables a direct implementation using the principles of

Kesner, F., Cimbalnik, J., Dolezalova, I., Brazdil, M. and Sekanina, L..

Fast Automated Interictal Spike Detection in iEEG/ECoG Recordings - {{\}it Using Optimized Memory Access}.

 2015 by SCITEPRESS – Science and Technology Publications, Lda. All rights reser ved

pipeline processing. It also offers a possibility to reg-

ulate the computational time needed for processing.

The following description will be devoted to a key

component of the algorithm - ﬁrst level detector. The

algorithm has been designed to be cache-effective in

order to achieve a high processing speed, which is

relatively rare in these types of algorithms (Ovchin-

nikov et al., 2010). Since an amount of data to be

processed is enormous, re-iterating through large data

arrays in memory, in which signals are stored, would

lead to a high cache miss ratio. In order to develop

a cache-effective algorithm, the algorithm is never re-

iterating through the signal array. All operations are

performed on signal data, when they come, sample

by sample. This approach was also used in order to

provide the real-time processing capability, which is

discussed later in this paper.

When the ﬁrst signal sample comes in, it is ﬁl-

tered by a ’spike band’ ﬁlter group (band pass 20-50

Hz) providing one spike band sample. This ﬁltered

sample may then be stored in a rotational buffer of

speciﬁed size (by parameter) or just used for com-

putation of adaptive min-max difference and then

dropped. Then the same signal sample (already stored

in cache/register) is ﬁltered again, this time by the sec-

ond ﬁlter group (band pass 1-35 Hz) providing one

’enhanced band’ signal sample. This ﬁltered sample

has to be stored in the rotational buffer of speciﬁed

size (by parameter) if the second level detection is ac-

tivated. Rotational buffers are computationally much

more effective than array shifting. They also increase

the cache-hit ratio and decrease memory requirements

compared to the block-processing oriented approach.

Compared to Barkmeier’s algorithm (Barkmeier

et al., 2012b), which uses the same frequency ﬁl-

tration bands, in this case, only one-directional ﬁl-

ters have been used instead of Matlab filtfilt()

function, which ﬁlters the signal from the left to the

right and subsequently from the right to the left in or-

der to eliminate phase delay, caused by ﬁltration.

In our algorithm, phase delay caused by ﬁlters is

eliminated after the ﬁnal detection step by subtracting

an empirical constant from the ﬁnal top and border

indexes of detected spike.

After each ﬁltered sample is obtained by ’spike

band’ ﬁlters, recent minimum and recent maximum

are estimated. Recent, or in other words, ’adaptive’

in this case means, that it represents the maximum or

minimum value in recently processed signal samples

(not associated with the rotational buffer size in any

way). This ’recentness’ is acquired by decay of these

values. The intensity of decay is one of the parameters

for ﬁrst level detection and it should be chosen based

on the sampling rate and amplitude scale of the signal.

After every sample is processed, these ’recent’ values

are diminished by decay parameter, either by subtrac-

tion or multiplication by parameter ≤ 1.0. Then the

recent minimum-maximum difference is computed.

When this min-max difference rises over a speciﬁed

threshold, then it is recognized as a rising edge event

and detected by the ﬁrst level method. Falling edge is

ignored. The illustration can be seen in Figure 1.

In order to avoid overlapping detections, possibly

of the same spike, the parameter of minimal distance

between rising edges is used.

Figure 1: Illustration of ﬁrst level detection method: B2 is a

regular signal channel with two detected spikes; ’enhanced

band’ is ﬁltered by bandpass 1-35 Hz; ’spike band’ is ﬁl-

tered by bandpass 20-50 Hz; min-max rising edge illustrates

threshold crossing points.

4 RESULTS

Papers dealing with spike detection methods don’t

often mention real computational time requirements

for signal processing (Lodder et al., 2013)(Barkmeier

et al., 2012b). However, it is one of the most essen-

tial parameters for spike detection algorithm evalua-

tion, when huge recording ﬁles are being processed.

Computational time results of our ﬁrst level detection

algorithm can be seen in Table 1. Measurements have

been done on HP Z420 workstation equipped with In-

tel(R) Xeon(R) CPU E5-1620.

Table 1: Algorithm speed / computational time measured on

35 minute ﬁle recorded with 5 kHz sampling rate consisting

of 150 channels. Data loading time from harddrive is not

included in the computing time, as these operations can be

overlapped.

CPU threads proc. time [s] sig./comp. t.

1 50.722 37.426

2 55.757 34.046

4 50.019 37.952

8 28.798 65.918

data loading 25.239 ← for illustration

4.1 Detection Sensitivity and Precision

A golden standard for spike detection practically does

not exist. As has been shown (Barkmeier et al.,

2012b), the inter-reviewer variability is huge. In or-

der to at least partially suppress this inter-reviewer

variability and also overlooking spikes, signals were

reviewed by a two member group of biomedical en-

gineers and only detections where full agreement was

achieved have been considered. Detections made by

this algorithm have been visualized into the signal

window by half-transparent marks. The group of re-

viewers has been counting missed spikes into one cat-

egory (false negatives), spike-free detections into an-

other (false positives) and correct detections (true pos-

itives) separately. The number of true negatives would

be hard to estimate because it is the rest of the signal

without marks. Computed precision and sensitivity

based on the preliminary evaluation are presented in

Table 2. According to these results, the algorithm per-

forms quite well, reaching sensitivity 86 - 97 % and

precision 95-99 %.

5 DISCUSSION

Original Barkmeier’s algorithm (Barkmeier et al.,

2012b) running in Matlab, takes on average 32 min-

Table 2: First level detector sensitivity and precision based

on preliminary testing over 3 different patients, where fre-

quently spiking channels consisting of 4x 35 minutes have

been in or adjacent to seizure onset zone, and rarely spiking

channels consisting of 15 x 35 minutes have been outside of

this area.

signal sensitivity [%] precision [%]

rare. spik. chs. 86.463 95.652

freq. spik. chs. 97.831 99.628

utes to process the 35 minutes long ﬁle. Our highly

optimized re-implementation of Barkmeier’s algo-

rithm in C, which has been created for fairness of

comparison, is approximately 8 times faster. Com-

pared to these implementations, this new ﬁrst level

detection algorithm implemented in C requires on av-

erage only 50 seconds running on the same hardware.

5.1 Real-time Detection

The algorithm has been designed with capability of

running the real-time detection. Hence this ﬁrst level

detector does not see practically any ’future’ signal

samples to perform a successful detection. The pos-

sible second level detection, that will be developed,

in order to compute spike features may need about

40-80 ms of ’future’ signal after the ﬁrst level detec-

tion. The algorithm has also been designed to keep the

computational time requirements minimal. In order

to illustrate how much computational time is needed

for a piece of signal, the signal-time / computational-

time ratio is presented in Table 1. It can be seen that

even without employing parallel computing for a 150

channel ﬁle with 5 kHz sampling rate, about 37-times

more channels can be processed while still perform-

ing in real-time. In theory it resulted in real-time pro-

cessing at 5550 channels.

ACKNOWLEDGEMENTS

This work was supported by Brno University of Tech-

nology grant under number FIT-S-14-2297 and by the

European Regional Development Fund the FNUSA-

ICRC project (No. CZ.1.05/1.1.00/02.0123).

REFERENCES

Barkmeier, D., Senador, D., Leclercq, K., and et al. (2012a).

Electrical, molecular and behavioral effects of inter-

ictal spiking in the rat. Neurobiology of Disease,

47(1):92–101.

Barkmeier, D., Shah, A., Flanagan, D., and et al.

(2012b). High inter-reviewer variability of spike de-

tection on intracranial eeg addressed by an automated

multi-channel algorithm. Clinical Neurophysiology,

123(6):1088–1095.

Lodder, S., Askamp, J., and van Putten, J. (2013). Inter-ictal

spike detection using a database of smart templates.

Clinical Neurophysiology, 124(12):2328–2335.

Ovchinnikov, A., Luttjohann, A., Hramov, A., and van Lui-

jtelaar, G. (2010). An algorithm for real-time detec-

tion of spike-wave discharges in rodents. Journal of

Neuroscience Methods, 194(1):172–178.

Staley, K., White, A., and Dudek, F. (2011). Interictal

spikes: harbingers or causes of epilepsy? Neuro-

science letters, 497(3):247–250.