Seizure Time Series Analysis I: Seizure Detection, Optimization and Assessment of Seizure Detection Algorithms

1
Seizure Time Series Analysis ISeizure
Detection, Optimization and Assessment of Seizure
Detection Algorithms

• Sridhar Sunderam, Ph.D.
• Center for Neural Engineering
• The Pennsylvania State University
• 4th Intl. Workshop on
• Seizure Prediction
• Kansas City, MO
• June 4, 2009

2
1. What to detect, and whyTargets
Fast ripples or HFOs 80-500 Hz

• Markers of epilepsy
• Interictal spikes
• Ripples, Fast ripples
• High frequency oscillations
• Electrographic seizures
• Clinical seizures
• Seizure precursor/Preictal state
• States of vigilance (?)

Staba et al., 2002
Clinical seizure
Andrzejak et al
3
1. What to detect, and whyApplications

• Treatment Goals
• Seizure control
• Improved QOL

Side-effects Cognition Allergies Interactions
Targets Ion channels Receptors
Antiepileptic drugs

• Uses of detection
• Diagnose seizure types and foci
• Evaluate for surgery
• Seizure warning
• Responsive therapy
• Treatment evaluation

• Surgery
• Localization
• Function mapping

Electrical stimulation
4
1. What to detect, and why Performance goals

• All of the above
• Early detection
• Sensitivity frequent hits
• Specificity rare misses
• Low false alarm rate patient anxiety, treatment
  dose
• Low cost (computation/power) implanted devices
• Quantitative description
• Not just onset intensity, duration, spread,
  dynamics

5
2. Seizure detectionThe jungle out there
Wavelet Based Automatic Seizure Detection in
Intracerebral EEG Identification of Ictal and
Pre-Ictal States Using RBF Networks with
Wavelet-Decomposed EEG Comparison of Seizure
Detection Algorithms in Pediatric
Patients Gaussian Process Modeling of EEG for the
Detection of Neonatal Seizures Detection of
Temporal Lobe Seizures from Scalp EEG
Recordings Epileptic Seizure Detection Using
Genetically Programmed Artificial
Features Automatic Detection of Spike and Wave
Discharges in the EEG of Genetic Absence Epilepsy
Rats Mixed-Band Wavelet-Chaos-Neural Network
Methodology for Epilepsy
Any sufficiently advanced technology is
indistinguishable from magic - Arthur C.
Clarke
6
2. Seizure detectionWhat you really want
Wavelet Based Automatic Seizure Detection in
Intracerebral EEG Identification of Ictal and
Pre-Ictal States Using RBF Networks with
Wavelet-Decomposed EEG Comparison of Seizure
Detection Algorithms in Pediatric
Patients Gaussian Process Modeling of EEG for the
Detection of Neonatal Seizures Detection of
Temporal Lobe Seizures from Scalp EEG
Recordings Epileptic Seizure Detection Using
Genetically Programmed Artificial
Features Automatic Detection of Spike and Wave
Discharges in the EEG of Genetic Absence Epilepsy
Rats Mixed-Band Wavelet-Chaos-Neural Network
Methodology for Epilepsy

• Signal source
• Cohort

• Focus/semiology
• Performance

7
2. Seizure detectionGeneral Framework
SIGNAL CONDITIONING

• TASKS
• Source selection
• Feature selection
• Filter design
• Detector design
• Performance evaluation

FEATURE EXTRACTION
Source
SEIZURE FILTERING
SEIZURE!
THRESHOLD
DETECTION
POST-PROCESSING
8
2. Seizure detectionSignal conditioning
(Topic of Litt lecture)
SIGNAL CONDITIONING
Signal

• Amplification
• Antialiasing/bandpass (e.g., 0.5-80 Hz)
• Sampling (e.g., 200 Hz)
• Artifact rejection
• Line noise, motion, stim, etc.

Here comes the good data!
Elimination of line noise
Gotman EEG Clin Neurophysiol 1982
9
2. Seizure detectionFeature extraction
Analysis window
Good data
FEATURE EXTRACTION

• Example features
• Wave morphology
• Amplitude
• Shape
• Spectral characteristics
• Band power
• Edge frequency
• Statistics
• Rhythmicity
• Entropy

One or more features that look different during
seizure
Half wave geometry Gotman EEG Clin Neurophysiol
1982
Line length feature Esteller et al. IEEE-EMBC 2001
10
2. Seizure detectionSeizure filtering
Wavelet-based FIR filters Osorio et al, 1998-2007
Brain chirps spectrographic signatures of
epileptic seizures Schiff et al., Clin
Neurophysiol 2000
Features
SEIZURE FILTERING
Correlates of Seizure content
11
2. Seizure detectionSeizure filtering
Example 1 Wavelet-based FIR filters Osorio et
al, 1998-2007
x(t)

b(t)
x
B(f)
X(f)
Features
SEIZURE FILTERING
Correlate of Seizure content
12
2. Seizure detectionSeizure filtering
Example 2 Spectrographic signatures of epileptic
seizures Schiff et al., Clin Neurophysiol 2000
Stacked in sequential windows
Brain chirps
Features
Correlation with seizure chirp template
SEIZURE FILTERING
Correlate of Seizure content
13
2. Seizure detectionPost-processing
Meng et al. Med Eng Phys 2004
Median filtering
Osorio et al. Epilepsia 1998
Background estimation
Gotman EEG Clin Neurophysiol 1982
Seizure content
POST-PROCESSING
Well-behaved output
14
2. Seizure detectionDetection/Classification
Decision Threshold
UNIVARIATE
Decision Boundary
AWAKE
SEIZURE!
SEIZURE
DETECTION
SLEEP
MULTIVARIATE (and multiclass)
Smooth output
15
2. Seizure detectionCase Study
Correlates of Seizure content
Chan et al. Clin Neurophysiol 2008
16
3. Seizure QuantificationWhy Quantify?

• Detection gives binary output
• Is there a seizure (onset)? (Y/N)
• There are seizures, and there are seizures
• Finer distinctions may be useful
• Treatment evaluation
• Measures of intensity, duration, spread
• Seizure dynamics
• Mechanism of initiation, progression or
  generalization

17
3. Seizure QuantificationUsing SDA output

• Already tracking seizure content
  just use it
• Combine measures to quantify relative
  severity
• Caveat Must capture the seizure, whole seizure,
  and nothing but!

Spread
Intensity
Duration
Osorio et al. Epilepsia 1998
18
3. Seizure QuantificationTime-frequency-energy
analysis
(topic of Franaszczuk lecture)
Jouny et al, Clin Neurophysiol 2003
19
4. SDA AssessmentGround Truth

• Human expert scoring is the gold standard
• Reproducibility, inter-rater reliability less
  than perfect

Four experts score one event
Wilson et al, Clin Neurophysiol 2003
20
4. SDA AssessmentPerformance Measures

• Single event
• Onset delay
• Offset time
• Area of spread

Spread

• Multiple events

Delay
SDA output
FP
TP
TP
TP
TN
FN

• Error rates
• Sensitivity TP/(TPFN)
• Specificity TN/(FPTN)
• Positive prediction value TP/(TPFP)

Clustering events affects error rates
21
4. SDA AssessmentROC Analysis
Detection threshold

• Assess performance
• Optimize params for
• Early detection
• Quantification
• Specific event type
• Individual/cohort

Non-seizure
Seizure
SDA output ?

• TPF Sensitivity
• FPF 1-Specificity
• AUC area under curve performance

http//www.anaesthetist.com/mnm/stats/roc/Findex.h
tm
22
4. SDA AssessmentOptimization
1. Parameter sensitivity analysis
2. Seizure filter adaptation
Haas et al. Med Eng Phys 2007
Osorio et al. Epilepsia 1998
23
5. Practical issuesArtifacts

• Line noise, motion, saturation, drop-out,
  cross-talk
• Mask activity, corrupt background
• Contribute to false detections
• Cause information loss
• Conditioning must flag/remove artifact
• Detector must disregard artifact
• Artifacts must not corrupt assessment

Chewing
Eye-blink
Saab Gotman Clin Neurophysiol 2005
Stimulation
Sun et al. Neurotherapeutics 2008
24
5. Practical issuesWhat is a True Negative?

• Specificity Fraction of non-seizures avoided
  (some use PPV, or FP/hr)
• But what is a True Negative?
• Whole interval between seizures? But duration
  varies
• Interictal epochs seizure duration? But how
  selected?
• Only epochs with interictal activity? Stringent
  but fair?
• Get a superset of detections by relaxing
  constraints

SDA output
Seizure
IED
25
5. Practical issuesProblem with FPR

• FPR FP per hour
• A common index of SDA performance
• FPR is a practical measure BUT
• FPR ? 1-Specificity
• FPR does not reflect seizure rate
• Best used in addition to
• Sensitivity TP/(TPFN)
• Specificity TN/(FPTN)
• PPV TP/(TPFP)

26
5. Practical issuesDealing with nonstationarity
Purves et al. 2001

• EEG changes with state of vigilance
• Therefore, SDA baseline is a moving target

27
5. Practical issuesDealing with nonstationarity
Malow et al. Epilepsia 1998

• Interictal spiking increases with sleep depth in
  temporal lobe epilepsy
• State of vigilance monitoring would be useful

28
5. Practical issuesThe price of failure

• Unwarranted anxiety and treatment
• False alarms, triggered stimulations
• Flawed treatment evaluation
• Altered study design e.g., closed-loop to
  open-loop stim
• Altered statistical power e.g., retrospective
  inclusion of FNs

TP FP TP FP TP TP
FP TP FP TP
Before TP FN TP FN TP
TP FN TP FN TP
After TP TP TP TP TP TP
TP TP TP TP
29
ConclusionWhich SDA is the best?
Or at least good enough

• The bare necessities
• Precedes clinical onset
• Enough lead time
• High specificity
• Why?
• To avoid cognitive impairment, and
• We dont really know when seizures start
• But treating subclinical events may be
  beneficial
• Ultimately
• What to detect, How, and Why
• IS REALLY UP TO YOU!
• THANK YOU