Epileptic Seizure Detection System

1
Epileptic Seizure Detection System

• Team Members
• Valerie Kuzmick, Biomedical Engineering
• John Lafferty, Computer Engineering
• April Serfass, Biomedical Engineering
• Doug Szperka, Computer Engineering
• Benjamin Zale, Computer Engineering

• Advisors
• Prawat Nagvajara, PhD, Computer Engineering
• Karen Moxon, PhD, Biomedical Engineering
• Jeremy Johnson, PhD, MCS/ECE

2
Problem Epilepsy

• Chronic Brain Function Disorder
• Characterized by Seizures
• Over two million suffering from epilepsy
• 1 of US population
• Current Treatments NOT Effective for 20 (400,000
  patients) of Epileptics

3
VISIONComplete System
Data AcquisitionSystem
Seizure Detection Unit
Stimulation Device
4
Design Challenge
Data AcquisitionSystem
Seizure Detection Unit
Stimulation Device
5
Prevention of Seizures

• NCP Brain Pacemaker
• Intermittent electrical pulses 24 hours a day
• Implanted under the collarbone
• Delivers electrical signals to the brain via
  vagus nerve in the neck
• When patient senses seizure coming, he or she can
  activate the stimulator manually

6
Developed Solution

• Prototype
• Microprocessor-based device that detects the
  neural activity associated with an epileptic
  seizure
• Results
• Seizure Detection 100 Accuracy
• Low False Positive Rate

7
Solutions for Seizure Detection

• Analysis of EEG Data With ANN
• Advantages
• Noninvasive
• Disadvantages
• Signal detection far from epicenter of seizure
• Loss of signal fidelity through bone scalp
• 65 detection rate

• Analysis of Multiple Single-Neuron Data
• Disadvantages
• Invasive
• Advantages
• Signal detection at the epicenter of seizure
• Ideal signal fidelity via direct recording from
  neurons
• Preliminary data suggest 100 detection rate

8
Method of Solution

• Data Collection Analysis
• Algorithm Development
• Software Simulation
• Detection Unit Implementation

9
Data Collection

• Certified laboratory rat handlers
• IACUC approved protocol
• Electrodes surgically implanted
• Temporal lobes
• PTZ administration
• Seizures induced

10
Data Collection
EIGHT-ARRAY ELECTRODE
RECORDING DEVICE
TEMPORAL LOBE
11
Multiple Single Neurons
12
Analysis

• Videotape
• Seizure/No Seizure
• NEX (NeuroExplorer)
• Rate Histograms
• Bin Size/Smooth Data
• Excel
• Imported NEX Files
• Seizures Distinguished
• Consolidation for Algorithm Development

13
Analysis
14
Algorithm Development

• Research from EEG Seizure Detectors
• Artificial Neural Network (ANN)
• Signal Processing Techniques
• Artificial Neural Network
• MATLAB Toolkit
• Created Various Feedforward Neural Networks
• Highest detection rate was 60

15
Cross Correlation Solution

• Neural activity becomes synchronized during a
  seizure
• Cross correlate data over a window of time
• Shows synchronization of neural action potentials
• Graphed the sum of pair-wise cross correlation
• Shape of the cross-correlation is determining
  factor

16
Data Conversion
17
Data Conversion
18
Cross Correlation Solution
19
Standard Deviation

• Statistic that tells you how tightly all the
  various data points are clustered around the mean
• Small standard deviation
• Data points are pretty tightly bunched together
• Large standard deviation
• Data points are spread apart

20
Cross Correlation Solution
Non Seizure Data
Seizure Data
21
Threshold Value

• Experimentally determined dividing line between
  seizure and non-seizure
• Algorithm Summary
• Data streamed into bins of finite length
• Cross Correlate
• Determine 1st standard deviation of cross
  correlated data
• Smaller than threshold value SEIZURE

22
Simulation

• Used MATLAB to Simulate
• Used Saved Data as Inputs
• Allowed Varying of Algorithm Parameters
• Saved Results of Each Run to File
• Final Parameters from Results
• Bin Size
• Bins per Window Size
• Threshold Value

23
Simulation Results

• 50ms Bin Size and 128 Bins per Window
• Promising Results
• Threshold Value was the Same
• Detected 100 of Observed Seizures
• Low False Positive Rate of 0.3 4.3 min/day
• Detected Seizures 4.5s Early on Average
• Some as early as 17s
• Few detected late 2.5s was the latest

24
Simulation Results
25
Detection Unit Implementation

• Implement algorithm to execute on dedicated
  microprocessor
• Speed
• Prototyping
• QED RM5231 RISC Processor
• MIPS Instruction Set
• V3 Hurricane Evaluation Board

26
Hardware

• Hurricane Evaluation Board
• Inserted into PCI slot of Windows-based computer
• Communication Protocols
• PCI
• Serial

27
Embedded Software

• ANSI C for portability
• Compiled into Motorola S-Record format
• Downloaded to board via serial port

28
Dataflow Diagram
 
 
 
 
 
 
Action Potential Data
NEX
Excel
RatStat (Hardware Simulation)
Data Concatenator
SerialComm
Hurricane Evaluation Board (Prototype)  
Simulation Output
Prototype Output
29
Host PC Software

• Automates Data Transmission
• Sums data into bins
• Generates S-Records of data
• Transmits data to evaluation board via serial
  port connection
• Tells evaluation board to execute embedded
  software
• Captures and reports seizure notification from
  evaluation board

30
Host PC Software
31
Economic Analysis

• Prototype Development
• Approximately 141,500 in equipment
• Future Commercial Development
• Needs to be System-on-a-Chip Solution
• Data Acquisition System 8,000
• Seizure Detection Unit 1,000
• NCP Brain Pacemaker 11,000
• Entire System 20,000 or less to be marketable
  and profitable

32
Results
Cross Correlation Window (bins) Cross Correlation Window (seconds) Average Execution Time (milliseconds)
32 1.6 13.2
64 3.2 50.3
128 6.4 182
256 12.8 718

• Prototype does not operate in real time when data
  is streamed

33
Conclusions

• Collected and Evaluated Approximately 1 Hour of
  Data from Three Specimens
• Only 45 minutes (2 Rats / 3 Trials) usable
• Remaining data corrupted
• 100 Seizure Detection Rate
• 0.3 False Positive Rate
• Seizures Predicted on an Average of 4.5 Seconds
  Beforehand

34
Automatic Seizure Detection System

• Team Members
• Valerie Kuzmick, Biomedical Engineering
• John Lafferty, Computer Engineering
• April Serfass, Biomedical Engineering
• Doug Szperka, Computer Engineering
• Benjamin Zale, Computer Engineering

35
Epileptic Episodes

• Encompasses Pre-Seizure and Seizure
• Highly correlated neural action potential data

36
Neural Action Potentials
37
Phase Angle Mapping
Results Indicate Seizure Detection Rate Greater
than 90
38
Frequency Content
Magnitude (dB)
Frequency (Hz)
39
Frequency Content
40
Phase Angle
41
Seizure Signature
42
Pattern Recognition
Weighted Sum of Action Potentials
Time (seconds)
43
Prototype
Data AcquisitionSystem
Seizure Detection Unit
Stimulation Device

• Receives Binary Data
• Processes Data Using Custom Algorithm
• Detects and Outputs Results