Seizure prediction: any better than chance?

Area of Science:

• Computational Neuroscience
• Epilepsy Research
• Biomedical Signal Processing

Background:

• Evaluating epileptic seizure prediction algorithms requires comparison against null hypotheses.
• Existing methods include analytical performance estimates and seizure predictor surrogates.
• A need exists for more robust methods to validate predictive algorithms.

Purpose of the Study:

• To extend the Monte Carlo framework for seizure predictor surrogates.
• To introduce and evaluate alarm times surrogates for hypothesis testing.
• To compare the efficacy of alarm times surrogates against analytical estimates.

Main Methods:

• Construction of artificial seizure time sequences and predictors.
• Testing algorithms against various null hypotheses under controlled conditions.
• Utilizing both analytical performance estimates and alarm times surrogates for comparison.

Main Results:

• Alarm times surrogates offer greater flexibility in testing null hypotheses compared to analytical estimates.
• Both methods demonstrate high statistical power in identifying true predictive capabilities.
• Analytical estimates exhibit bias, leading to false positive rejections with long inter-alarm intervals for Poisson predictors.

Conclusions:

• Alarm times surrogates present significant advantages over traditional analytical performance estimates.
• This surrogate method aids in resolving whether seizure prediction algorithms possess genuine predictive power or are subject to chance.
• The findings are crucial for advancing the reliability of seizure prediction technologies.