Study shows how machine learning techniques reduce false alarms in epilepsy seizure prediction
Besides developing the transfer learning model, the RECoD project has also advanced unsupervised methods for detecting the pre-seizure state, as well as for interpreting and validating seizure prediction models
A study conducted by the Department of Informatics Engineering (DEI) of the Faculty of Sciences and Technology of the University of Coimbra (FCTUC) shows that transfer learning (TL) approaches lead to a reduction in the number of false alarms in seizure prediction while maintaining the same predictive capacity in machine learning (ML) models.
The research was carried out as part of the project “RECoD-Towards Realistic Epileptic Seizure Prediction: dealing with long-term concept drifts and data-labeling uncertainty,” and has been published in the journal “Scientific Reports.”
TL is an ML technique in which a model pretrained on one task is adapted to a new, related task. Training a new ML model is time-consuming and intensive, requiring large amounts of data, high computational cost, and multiple iterations before it is ready for production. In contrast, TL retrains existing models on related tasks, which are then fine-tuned with new data.
"This paper presents a transfer learning (TL) approach to develop patient-specific epileptic seizure prediction models based on deep neural networks (DNNs). The model was developed using data from 41 patients from the EPILEPSIAE database, which was then used to optimise personalised seizure prediction models," explains César Teixeira, Professor at DEI and researcher at the Centre for Informatics and Systems of the University of Coimbra (CISUC).
The author of the article and head of the project adds that the results “showed that the development of a transfer learning model allowed us to obtain about four times fewer false alarms while maintaining the same ability to predict crises as when the models were trained from scratch. Therefore, we can conclude that the limitations imposed by the small number of crises can be overcome using these techniques.”
According to the FCTUC professor, besides developing the transfer learning model, the RECoD project has also advanced unsupervised methods for detecting the pre-seizure state, as well as for interpreting and validating seizure prediction models. The consortium for this project included the FCTUC, the University Hospital of Freiburg (UKFR), and the Coimbra Hospital and University Centre (CHUC).
The scientific article “Addressing data limitations in seizure prediction through transfer learning” is available here.