The Seizure Prediction Contest

1
The Seizure Prediction Contest

• Andreas Schulze-Bonhage, Jean Gotman and Björn
  Schelter

2
Objective

• Several publications concerning algorithms
  designed for seizure prediction
• Usually retrospectively assessed
• In-sample testing
• Algorithms are applied to own data only, no
  evaluation on a common data pool

3
The Idea

• Fair assessment of seizure prediction performance
  based on learning and testing data only the
  learning data are available for optimization of
  prediction methods
• Parameters can be tuned based on these learning
  data
• Analysis on the testing data is done by the
  Freiburg group referee Jean Gotman

4
The data

• Selection of the data by Jean Gotman
• 3 patients selected
• Patients characteristics are

5
Patient 1
Age at invasive monitoring 30
Sex female
Diagnosis focal cortical dysplasia
Epilepsy surgery left temporolateral resection sparing the hippocampus
Outcome of epilepsy surgery engel Ia
Seizures recorded simple-partial, complex-partial and secondarily generalized tonic-clonic
Electrode types depth, strip
No. of channels 60
Sampling rate 512 Hz
6
Patient 1
7
Patient 2
Age at invasive monitoring 17
Sex male
Diagnosis left temporal tumor
Epilepsy surgery resection of tumor
Outcome of epilepsy surgery engel Ia
Seizures recorded simple-partial, complex-partial and secondarily generalized tonic-clonic
Electrode types depth, strip
No. of channels 44
Sampling rate 512 Hz
8
Patient 2
9
Patient 3
Age at invasive monitoring 10
Sex male
Diagnosis hippocampal sclerosis, mesiotemporal focal cortical dysplasia and an additional right occipital lesion of unknown histology
Epilepsy surgery selective right amygdalo-hippocampectomy
Outcome of epilepsy surgery engel Ia
Seizures recorded simple-partial, complex-partial and secondarily generalized tonic-clonic
Electrode types depth, strip
No. of channels 22
Sampling rate 512 Hz
10
Patient 3
11
The learning data Pat 1
12
The learning data Pat 2
13
The learning data Pat 3
14
The participants
More than 50 groups analyzed (?) the data
15
The participants
More than 50 groups downloaded (?) the data
16
The participants
More than 50 groups registered for the contest
17
The participants
Name Affiliation
Levin Kuhlmann, Elma O'Sullivan-Greene, Andrea Varsavsky, Craig Savage University of Melbourne, Australia
Stephan Bialonski, Anton Chernihovskyi, Dieter Krug, Jens Prusseit, Matthäus Staniek, Florian Mormann, Klaus Lehnertz Department of Epileptology, University of Bonn, Germany
These two groups submitted three algorithms.
18
The participants
Name Affiliation
Pooja Rajdev Purdue University
Tay Netoff University of Minnesota
Vadim Smolyakov
These three groups submitted three algorithms.
19
The results
Learning data Learning data Learning data Learning data Learning data Learning data Testing data Testing data Testing data
Fraction Fraction Fraction Fraction
Pooja True Pred False Pred Num Seiz Sens Dur. corr Dur. Incorr Sens Dur. corr Dur. Incorr
Pat 1 3 258 6 50 0 20 0 0 20
Pat 2 2 99 10 20 0 20 20 0 20
Pat 3 2 175 5 20 40 20 0 0 0

Vadim
Pat 2 3 10 10 30 0 20 70 0 70

Tay
Pat 2 10 3040 10 100 20 50 100 0 90
Pat 3 5 6714 5 100 0 30 80 10 70
20
The results
Learning data Learning data Learning data Learning data Learning data Learning data Testing data Testing data Testing data
Fraction Fraction Fraction Fraction
Pooja True Pred False Pred Num Seiz Sens Dur. corr Dur. Incorr Sens Dur. corr Dur. Incorr
Pat 1 3 258 6 50 0 20 0 0 20
Pat 2 2 99 10 20 0 20 20 0 20
Pat 3 2 175 5 20 40 20 0 0 0

Vadim
Pat 2 3 10 10 30 0 20 70 0 70

Tay
Pat 2 10 3040 10 100 20 50 100 0 90
Pat 3 5 6714 5 100 0 30 80 10 70
THANKS To all 3 participants who submitted an
algorithm
21
The contest continues

• Several groups are interest in the contest
• The main problem has been time limitation
• Very important and timely question
• No clear-cut winner so far
• The contest continues until the next workshop in
  approximately 1.5 years

22
The contest continues

• Several groups are interest in the contest
• The main problem has been time limitation
• Very important and timely question
• No clear-cut winner so far
• The contest continues but
• from now on you have to pay to
• participate in the contest