2. 2Challenge the future
Epilepsy
• Epileptic seizure
• Signs/Symptoms
• Synchronous neuronal activity
• Absence seizures
• Type of epileptic seizure
• Brief loss of consciousness
Definition:
“A brain disorder characterized by epileptic seizures”
Video of absence seizure. Go to:
https://www.youtube.com/watch
?v=obbg1BFt26Q
3. 3Challenge the future
How is epilepsy treated?
• Conventional treatment methods
• Drugs
• Resective Surgery (severe cases)
• New technique emerging
• Neurostimulation Electrical / Optogenetic
• One FDA approved device (Neuropace RNS)
• Problem statement
Can we detect seizures fast and reliably using a
seizure-detection system which is suitable for
implantable (ultra-low power) application?
4. 4Challenge the future
Objectives of this work: Outline
• Implementation of a detection mechanism
• Prototype realization
• Evaluation of mechanism
• Detection performance
• Computational overheads
• Trade-offs
7. 7Challenge the future
Implementation:
Complex Morlet wavelet
• Selection reasons
• Similar morphology
• Time-frequency localization
• Scalogram
• Average wavelet output of all
seizure intervals
• Power of filter at different
frequency scales
• Onset at 1 second
• Max power at 7Hz
• Implement FIR
8. 8Challenge the future
• FIR filter coefficients
• Sampled at 100Hz
• Total of 229
• Filter output:
• Power of filter output y(n)
to distinguish a seizure
Implementation:
Finite Impulse Response
25. 25Challenge the future
Conclusions
• Range of possible solutions
• Sens=98%, Delay=380ms, Spec=87%
• Sens=70%, Delay=850ms, Spec=98%
• Detection delay of 492ms (Adaptive threshold)
• 970ms (fastest solution available)
• Trade-off
• Detection accuracy vs. Energy consumption
• Real-time constraints met on SiMS
Can we detect seizures fast and reliably using a
seizure-detection system which is suitable for
implantable (ultra-low power) application?
26. 26Challenge the future
Conclusions
YES WE CAN
Can we detect seizures fast and reliably using a
seizure-detection system which is suitable for
implantable (ultra-low power) application?
27. 27Challenge the future
Future work & uStim
• Interesting extensions
• Other wavelets
• Improve adaptive mechanism
• ASIC implementation of solution
• SiMS with application specific units ( e.g. multiplier, SiMD unit)
• Part of this work
• Prototype for tinnitus
treatment
• Electrical Neurostimulation
Transient occurrence of signs caused excessive synchronous neuronal brain activity
thesis focus on detection and suppression of absence
purpose specialized/effective treatment
Optogenetic stimulation light is applied on genetically sensitized neurons
FDA -> 50 % seizure reduction
What are requirements to build such systems?
In order to answer in the next slides I am going to describe the step that were taken.
How does a seizure look like?
Electrocorticogram
Signals that come out of the brain
Snapshot of an ECoG recording taken from a mouseSpike-Wave Discharges repeat at in frequency range 6-8Hz
In order to detect and suppress seizures we created a closed loop system
Which consists of
Analyze component ( input signals, Analog, Digital ( 100Hz, feature extraction, decision making) , stimulator )
Optogenetic stimulation , applies light to genetically sensitized neurons
Focus of our work on digital
What it does? Used to distinguish seizures from non-seizures
Why?
Localized time and frequency, which gives frequency information on the signal in time domain
Compared to other time-frequency the wavelet can provide better resolution by selecting an appropriate wavelet. (in this case the complex Morlet)
How?
Scalogram by averaging the filter output of seizure interval we obtain
Present the power of the filter
Align onset @ 1 sec
Note increase in power certain frequency scale & only after the seizure-onset.
Practical FIR filter implementation the impulse response of continuous wavelet filter sampled at 100Hzand using the effective support in matlab we obtained 229 coefficient for each of the complex components
Resulting impulse response is seen in figure
Filter is calculated by the convolution of the impulse response and the input signal
From the scalogram we saw that power is good indicator to distinguish a seizure interval . The sum of the square values of imaginary and real part
Increased output during seizures
Adaptive threshold will be defined later
First to verify that the system works, we did in-vivo experiments.
Actually it was required because they needed to do the experiments.
This also helped us to prove the concept.
Built our own PCB for the analog module & ported the detection system to a Beaglebone platform
It works. See example where SWD stop
Preliminary wavelet filter analysis to reduce filter response. However
I won't tell you the details(which can be found in thesis) since we will talk about a very extensive filter analysis in just a few minutes...
Moving on of the evaluation of closed-loop system
In our evalulationThe dataset we used to evaluate the detection performance of ….
In the golden standard the seizures are defined by SWD activity that lasts more that 1 second see pictureAlso considered an extra case were we have at least 3 SWD. However tNot presented due to timing limitations and can be found in thesis
Is evaluated on
1.Coefficient window
present the results on different size effectively filter order , also offset details can be found in thesis2. Different threhold first static ( fair comparison between filter relative to its filter instance output range, don’t want to compare absolute values
and then adaptive
Explain why:
Window size affect output range threshold levels are relative to filter output therefore no effect
Higher threshold missing seizure with lesser magnitude filter output
Explain why
Higher threshold less FP Larger window better filter response to seizure
Mention that lowering thresh reduces detection delay
Explain whyLarger thesh higher delay
Larger window, delay filter response
!!Blank page!!!In thesis you can also find the Pareto optimal points of all three performance metrics are computed
Describe how tup and tdown affects
ADR describes the average detection performance
For a window size of 64
Explain why better in detail
Line-length : not finely tuned as wavelet filter
Daubechies wavelet : Similar, extra step where they Average wavelet filter output which introduces delay better results though
ANN : compex artificial neural network adds delay. Slightly better specificity
Advantage of our solution : simple, and from our evaluation small training set
Focus on execution to see if implementation is viable/feasible
Energy consumption to evaluate the costs of realizing this solution
Battery lifespan indication of how long can such a system last
Window size is enough 64 fof adaptive
Linear increases. Almost doubles evety window size doubling
For instance on a battery 0.1Wh hour 8 coefficients 7.8 weeks ( After that recharging, solution wireless charging )
Combination detection performance and computational overheads
ADR as an indication of performance