Zoom out Search Issue

ManualsBrandsContents Manualsaudio & home theatreZoom in

IEEE SIGNAL PROCESSING MAGAZINE [14] MARCH 2015

[

special

REPORTS

]

continued

such a nerve might contain hundreds or

thousands of individual fibers [axons],

with signals propagating over a wide

range of velocities (up to 120 m/s in hu-

mans) in both directions. “This (tech-

nique) should be useful because neural

propagation velocity and fiber diameter

are generally related, so an analysis of

activity by velocity and direction is

equivalent to knowing the diameters of

the nerves that are excited at that time,”

Taylor says. “Anatomy then allows us to

link this to function.”

In the future, Taylor says these func-

tion-specific signals might be used to de-

sign systems for controlling neuropros-

thetic devices, such as providing a neural

stimulator with a feedback loop for blad-

der control to treat urinary incontinence.

“Currently available methods to provide

the information we seek tend to rely on

fairly classical pattern processing meth-

ods, such as clustering and principal com-

ponent analysis under the generic title of

‘spike sorting’,” Taylor says. Such methods

tend to be computationally intensive and

therefore unattractive for implantation.

“In addition, some form of training is gen-

erally required and may be impossible or

impracticable,” he adds. By contrast the

signal processing required for VSR is com-

putationally simple, power efficient and

lends itself to real-time working.

Taylor says that signal processing is a

key building block in the group’s research.

“This is because surgical considerations

impose strict limits on the size and com-

plexity of our implanted devices and hence

on the sensitivity and resolution of our ba-

sic signal acquisition capability,” he ex-

plains. “Signal processing can compensate

for this and is used wherever possible for

filtering noise, performing spectral analy-

sis of waveforms, and ultimately for decod-

ing the impulses that we record from the

nervous system.”

According to Taylor, VSR requires

multiple samples of the composite prop-

agating neural signal. Such samples are

typically provided by a multielectrode

cuff (MEC) placed around the nerve

(Figure 2). The MEC, which is an insu-

lating cuff typically 2–3 cm in length

and containing 10–12 electrodes, is an

extension of the traditional tripolar type

of nerve cuff that has been implanted in

many patients successfully for several

decades, Taylor says.

The samples are identical but delayed

by a period that depends on both the cuff

geometry and the propagation velocity of

the signal. To construct the velocity spec-

trum from this data an operation called

“delay-and-add” is applied. The operation

adds artificial delays that cancel the natu-

ral delays in each channel before finally

adding all the signals together. “When the

artificial delays are equal to the naturally-

occurring ones, the spectral output passes

through a peak (local maximum) indicat-

ing the presence of an excited population

of axons at that velocity,” Taylor says. “This

is the simplest approach to VSR and the

resulting spectrum is called the intrinsic

velocity spectrum (IVS).” The method, he

notes, is closely related to various beam-

forming algorithms employed in radio and

radar antenna systems.

Unfortunately the method achieves rel-

atively poor velocity selectivity, Taylor

says. It has particular difficulty in distin-

guishing closely spaced velocity peaks.

Various additional techniques have been

developed to improve the velocity selectiv-

ity including the use of bandpass filters

and time delay neural networks (TDNNs),

Taylor explains.

One of the biggest limitations inherent

in existing neural signal processors is the

requirement to build complex statistical

models. “These models are not only com-

putationally expensive to produce but also

require a good deal of time to ‘learn’ as

they become patient-specific,” Taylor says.

“To overcome these limitations we consid-

ered an entirely different signal processing

approach, based on conduction velocity

instead of pattern shape.”

Noise poses another challenge. “The

signals we record are from biological

sources and so are often very noisy,” Tay-

lor remarks. “It is not uncommon for the

signal-to-noise ratio (SNR) to be less than

0 dB, and so innovative methods must be

developed to extract information.” Cou-

pled with the requirements for real-time

operation and good long-term stability,

the challenges are not insignificant.

Recording neural activity from an in-

tact nerve represents another highly chal-

lenging task, due to the poorly understood

nature of the electrode-tissue interface

and the associated problems of handling

[FIG2] A multielectrode nerve cuff used for velocity-selective recordings made by Martin

Schuettler, a senior scientist at the University of Freiburg and chief technology officer

of CorTek, a Freiuburg, Germany-based developer of a neurotechnological platform for

measuring and stimulating of brain activity. (Photo credit: Martin Schuettler.)

THE WORLD’S NEWSSTAND

THE WORLD’S NEWSSTAND