Prediction of epileptic seizures using fNIRS and machine learning

Research to predict epileptic seizures has been mainly focused on the analysis of electroencephalography (EEG) signals; however, recent research efforts have encouraged the use of a relatively new optical signal modality, called functional Near-Infrared Spectroscopy (fNIRS). In fNIRS, near-infrared light is injected into the scalp and the intensity of the reflected light is registered in optodes. Light absorption in hemoglobin depends on the level of blood oxygenation, which is related to brain activity. In this technique, two parameters are measured at each optode, the relative level of oxygenated hemoglobin (HbO) and the relative level of deoxygenated hemoglobin (HbR). In this work we investigated the feasibility of predicting epileptic seizures, using either fNIRS, EEG, or a combination of both signals. In one set of experiments, different implementations for epileptic seizure prediction are tested by using (1) different combinations of electrical and optical signals (EEG, HbO, HbR, EEG+HbO, EEG+HbR, HbO+HbR, EEG+HbO+HbR) and (2) two different classifiers, (Support Vector Machine - SVM and Multi-Layer Perceptron - MLP). In the second set of experiments, seizures are predicted within a five-minute window that is moved up to 15 minutes before the start of the epileptic seizure. By computing the Positive Predictive Value (PPV) and the accuracy, it is demonstrated that fNIRS -based epileptic prediction outperforms EEG-based epileptic prediction. By using optical signals and the SVM classifier, a PPV greater than 99% and an accuracy of 100% were obtained. PPV values of 100% are also obtained when seizures are predicted up to 15 minutes in advance. Furthermore, Kernel Discriminant Analysis (KDA) is used to demonstrate that the highest separability among the classes, corresponding to different epileptic signal phases (pre-ictal, ictal, and inter-ictal), is achieved when fNIRS recordings are used as features for prediction. Finally, fNIRS-based epileptic seizure prediction is tested with Random Chance classifiers. In this study, we showed that fNIRS signals are an effective tool to predict epileptic seizures, even without the use of EEG signals, which are the current standard for seizure prediction.