Aura Algorithms
Epilepsy seizure prediction is a very wide challenge. In 2017, International League Against Epilepsy revised an expanded ~30 seizure types classification[1]. In addition, each patient physiological response to epilepsy seizure is unique.
Our approach is to learn incrementally to detect patient-specific patterns corresponding to seizure pre-ictal phase. To do so we will rely on data science recent development in Supervised learning algorithms. Recurrent neural networks and especially LSTM networks algorithms have shown outstanding performances on multiples applications such as speech recognition or automatic image captioning and suits very well to multi-label time series classification[2][3].
Step 1: Training Phase
For each patient, we will feed the algorithm with physiological data, reported auras and seizures to train it to classify patient pre-ictal seizure periods from regular physiological activity. Once we reach a good algorithm performance (cf Algorithm evaluation) we will initiate the monitoring phase.
Step 2: Monitoring Phase
The algorithm will process patient physiological data on the fly and triggers an alert when it detects a seizure pre-ictal phases.
Algorithm evaluation
We will use 3 metrics to assess the performance of epilepsy seizure prediction algorithms:
- seizure prediction sensitivity
- number of false prediction
- time interval between seizure prediction alert and seizure onset
Source Code
The Aura Algorithm source code are distributed under GPL v3 license and available on Github.
References
[1] Robert S. Fisher, J. Helen Cross, Jacqueline A. French, Norimichi Higurashi, Edouard Hirsch, Floor E. Jansen, Lieven Lagae, Solomon L. Moshe, Jukka Peltola, Eliane RouletPerez, Ingrid E. Scheffer, and Sameer M.Zuberiobert - Operational classification of seizure types by the International League Against Epilepsy: Position Paper of the ILAE Commission for Classification and Terminology. Epilepsia, 1–9, 2017
[2] Zachary C. Lipton, David C. Kale, Charles Elkan, Randall Wetzel, Laura P. and Leland K. Whittier - Learning to diagnose with LSTM reccurent neural networks - Conference paper at ICLR, 2016
[3] David Ahmedt-Aristizabal, Clinton Fookes, Kien Nguyen and Sridha Sridharan - Deep Classification of Epileptic Signals - ARXIV, Computer Science - Learning, Quantitative Biology - Neurons and Cognition, 2018