RESUMO
Epilepsy is regarded as a structural and functional network disorder, affecting around 50 million people worldwide. A correct disease diagnosis can lead to quicker medical action, preventing adverse effects. This paper reports the design of a classifier for epilepsy diagnosis in patients after a first ictal episode, using electroencephalogram (EEG) recordings. The dataset consists of resting-state EEG from 629 patients, of which 504 were retained for the study. The patient's cohort exists out of 291 patients with epilepsy and 213 patients with other pathologies. The data were split into two sets: 80% training set and 20% test set. The extracted features from EEG included functional connectivity measures, graph measures, band powers and brain asymmetry ratios. Feature reduction was performed, and the models were trained using Machine Learning (ML) techniques. The models' evaluation was performed with the area under the receiver operating characteristic curve (AUC). When focusing specifically on focal lesional epileptic patients, better results were obtained. This classification task was optimized using a 5-fold cross-validation, where SVM using PCA for feature reduction achieved an AUC of 0.730 ± 0.030. In the test set, the same model achieved 0.649 of AUC. The verified decrease is justified by the considerable diversity of pathologies in the cohort. An analysis of the selected features across tested models shows that functional connectivity and its graph measures have the most considerable predictive power, along with full-spectrum frequency-based features. To conclude, the proposed algorithms, with some refinement, can be of added value for doctors diagnosing epilepsy from EEG recordings after a suspected first seizure.
RESUMO
The detection of causal effects among simultaneous observations provides knowledge about the underlying network, and is a topic of interests in many scientific areas. Over the years different causality measures have been developed, each with their own advantages and disadvantages. However, an extensive evaluation study is missing. In this work we consider some of the best-known causality measures i.e., cross-correlation, (conditional) Granger causality index (CGCI), partial directed coherence (PDC), directed transfer function (DTF), and partial mutual information on mixed embedding (PMIME). To correct for noise-related spurious connections, each measure (except PMIME) is tested for statistical significance based on surrogate data. The performance of the causality metrics is evaluated on a set of simulation models with distinct characteristics, to assess how well they work in- as well as outside of their "comfort zone." PDC and DTF perform best on systems with frequency-specific connections, while PMIME is the only one able to detect non-linear interactions. The varying performance depending on the system characteristics warrants the use of multiple measures and comparing their results to avoid errors. Furthermore, lags between coupled variables are inherent to real-world systems and could hold essential information on the network dynamics. They are however often not taken into account and we lack proper tools to estimate them. We propose three new methods for lag estimation in multivariate time series, based on autoregressive modelling and information theory. One of the autoregressive methods and the one based on information theory were able to reliably identify the correct lag value in different simulated systems. However, only the latter was able to maintain its performance in the case of non-linear interactions. As a clinical application, the same methods are also applied on an intracranial recording of an epileptic seizure. The combined knowledge from the causality measures and insights from the simulations, on how these measures perform under different circumstances and when to use which one, allow us to recreate a plausible network of the seizure propagation that supports previous observations of desynchronisation and synchronisation during seizure progression. The lag estimation results show absence of a relationship between connectivity strength and estimated lag values, which contradicts the line of thinking in connectivity shaped by the neuron doctrine.
RESUMO
Neural desynchronization was shown as a key mechanism of vagus nerve stimulation (VNS) action in epilepsy, and EEG synchronization measures are explored as possible response biomarkers. Since brain functional organization in sleep shows different synchrony and network properties compared to wakefulness, we aimed to explore the effects of acute VNS on EEG-derived measures in the two different states of vigilance. EEG epochs were retrospectively analyzed from twenty-four VNS-treated epileptic patients (11 responders, 13 non-responders) in calm wakefulness and stage N2 sleep. Weighted Phase Lag Index (wPLI) was computed as connectivity measure of synchronization, for VNS OFF and VNS ON conditions. Global efficiency (GE) was computed as a network measure of integration. Ratios OFF/ON were obtained as desynchronization/de-integration index. Values were compared between responders and non-responders, and between EEG states. ROC curve and area-under-the-curve (AUC) analysis was performed for response classification. In responders, stronger VNS-induced theta desynchronization (p < 0.05) and decreased GE (p < 0.05) were found in sleep, but not in wakefulness. Theta sleep wPLI Ratio OFF/ON yielded an AUC of 0.825, and 79% accuracy as a response biomarker if a cut-off value is set at 1.05. Considering all patients, the VNS-induced GE decrease was significantly more important in sleep compared to awake EEG state (p < 0.01). In conclusion, stronger sleep EEG desynchronization in theta band distinguishes responders to VNS therapy from non-responders. VNS-induced reduction of network integration occurs significantly more in sleep than in wakefulness.