No abstract

Electroencephalographic (EEG) signal processing and machine learning can support neurologists’ work in discriminating Psychogenic Non-Epileptic Seizure (PNES) from epilepsy. PNES represents a neurological disease often misdiagnosed. Although the symptoms of PNES patients can be similar to those exhibited by epileptic patients, EEG signals during a psychogenic seizure do not show ictal patterns such as in epilepsy. Therefore, PNES diagnosis requires long-term EEG video. Applying signal processing and machine-learning methodologies could help clinicians find helpful information in the clinical diagnosis of PNES by analyzing EEG signals registered in resting conditions and in a short time. These methodologies should prevent long EEG recording sessions and avoid inducing seizures in the subjects. The aim of our study is to develop and validate several machine-learning models on a larger dataset, consisting of 225 EEGs (75 healthy, 75 PNES, and 75 subjects with epilepsy). A deep analysis of our results shows that changes in the evaluation strategy led to changes in accuracy from 45% to 83.98% for a standard Light Gradient Boosting Machine (LGBM) classifier. Our findings suggest that it is necessary to operate a very rigorous control in terms of experimental data collection (patient selection, signal acquisition) and terms of validation strategies to obtain and reproducible results.

No abstract