1. Multi-time scale modelling for improved seizure prediction
Levin Kuhlmann, PhD
Swinburne University of Technology and University of Melbourne

2. First-in-man prospective trial
15 patients
Continuous recording 6 months – 3 years
Prediction is possible
Did not work for everyone
Results were variable between patients
15 patients
No other dataset in the world that matches this level of detail (time and resolution)
Seizure prediction is possible – 10 people had above chance prediction, 5 were excellent

3. How do we get to the next prospective trial?
Find algorithms that maximize retrospective performance
How can we do this?
15 patients
No other dataset in the world that matches this level of detail (time and resolution)
Seizure prediction is possible – 10 people had above chance prediction, 5 were excellent

4. Not all seizures are the same
Cook et al., (2015) Epilepsia

5. Feature
Not all seizures are the same
Feature

6. Probabilistic model Trial Length (Days) Time of Day
Reduce to a single variable - time
Time of Day

7. Multi-time scale modelling for improved seizure prediction
Short-time scales: standard feature calculation (typically less than 60s windows)
Long-time scales:
Time of day (hourly)
Time of month
Season
How do we combine these?
15 patients
No other dataset in the world that matches this level of detail (time and resolution)
Seizure prediction is possible – 10 people had above chance prediction, 5 were excellent

8. Multi-time scale modelling for improved seizure prediction
Short-time scales: standard feature calculation (typically less than 60s windows)
Long-time scales:
Time of day (hourly)
Time of month
Season
How do we combine these? Condition short-time scale-based predictors on long-time scale variables.
p(seizure|variable)
15 patients
No other dataset in the world that matches this level of detail (time and resolution)
Seizure prediction is possible – 10 people had above chance prediction, 5 were excellent

9. https://academic.oup.com/brain/article-abstract/140/8/2169/4032453

10. Probabilistic model Categorical prediction Vs Probabilistic forecast
Test using a simple classifier

11. Probabilistic model Logistic regression seizure non-seizure
Test using a simple classifier

12. Probabilistic model Bayes rule Logistic regression Prior Likelihood
Re-write logistic regression model in terms of a prior probability
Prior
Likelihood

13. Probabilistic model Bayes rule Logistic regression
Using our time of day we end up with an adaptive threshold classifier
This is a first test. My hope is that this sort of calibration will become routine in implantable devices – it doesn’t matter what algorithm is being used as long as it is a generative (rather than discriminative) method.

14. Signal features
signal energy in four frequency bands (8–16, Hz, 16–32 Hz, 32–64 Hz, 64–128 Hz)
line length
5 metrics x 16 channels = 80 features > select 16
Test using a simple classifier

20. https://academic.oup.com/brain/article-abstract/140/8/2169/4032453

21. Conclusion
Seizure prediction conditioned on time of day or simple time of day prediction offer new ways to gain improvements
Probabilistic framework enables conditioning on other variables besides time:
Weather
Hormone levels
Arousal levels
Heart rate
Patient specificity and long-term monitoring essential
15 patients
No other dataset in the world that matches this level of detail (time and resolution)
Seizure prediction is possible – 10 people had above chance prediction, 5 were excellent

22. Acknowledgements
University of Melbourne Philippa Karoly David Grayden Dean Freestone Mark Cook
Collaborators
Hoameng Ung Kent Leyde

24. Seizure Patterns: Weather
Seizure Probability
Temperature (degrees C)
unpublished

27. Melbourne University AES-MathWorks-NIH Seizure Prediction Challenge
478 teams, 646 data scientists, algorithms
$20,000 prize money
closed on the 1st of Dec. 2016