The goal of explaining black boxes in EEG seizure prediction is not to explain models' decisions.

Many state-of-the-art methods for seizure prediction, using the electroencephalogram, are based on machine learning models that are black boxes, weakening the trust of clinicians in them for high-risk decisions. Seizure prediction concerns a multidimensional time-series problem that performs continuous sliding window analysis and classification. In this work, we make a critical review of which explanations increase trust in models' decisions for predicting seizures. We developed three machine learning methodologies to explore their explainability potential. These contain different levels of model transparency: a logistic regression, an ensemble of 15 support vector machines, and an ensemble of three convolutional neural networks. For each methodology, we evaluated quasi-prospectively the performance in 40 patients (testing data comprised 2055 hours and 104 seizures). We selected patients with good and poor performance to explain the models' decisions. Then, with grounded theory, we evaluated how these explanations helped specialists (data scientists and clinicians working in epilepsy) to understand the obtained model dynamics. We obtained four lessons for better communication between data scientists and clinicians. We found that the goal of explainability is not to explain the system's decisions but to improve the system itself. Model transparency is not the most significant factor in explaining a model decision for seizure prediction. Even when using intuitive and state-of-the-art features, it is hard to understand brain dynamics and their relationship with the developed models. We achieve an increase in understanding by developing, in parallel, several systems that explicitly deal with signal dynamics changes that help develop a complete problem formulation.

The impact of cathodal tDCS on the GABAergic system in the epileptogenic zone: A multimodal imaging study.

Objectives: We aimed to investigate the antiepileptic effects of cathodal transcranial direct current stimulation (c-tDCS) and mechanisms of action based on its effects on the neurotransmitters responsible for the abnormal synchrony patterns seen in pharmacoresistant epilepsy. This is the first study to test the impact of neurostimulation on epileptiform interictal discharges (IEDs) and to measure brain metabolites in the epileptogenic zone (EZ) and control regions simultaneously in patients with pharmacoresistant epilepsy. Methods: This is a hypothesis-driven pilot prospective single-blinded repeated measure design study in patients diagnosed with pharmacoresistant epilepsy of temporal lobe onset. We included seven patients who underwent two sessions of c-tDCS (sham followed by real). The real tDCS session was 20 min in duration and had a current intensity of 1.5 mA delivered via two surface electrodes that had dimensions of 3 × 4 cm. The cathode electrode was placed at FT7 in the center whereas the anode at Oz in the center. After each session, we performed electroencephalographic recording to count epileptiform IEDs over 30 min. We also performed magnetic resonance spectroscopy (MRS) to measure brain metabolite concentrations in the two areas of interest (EZ and occipital region), namely, gamma-aminobutyric acid (GABA), glutamate (Glx), and glutathione. We focused on a homogenous sample where the EZ and antiepileptic medications are shared among patients. Results: Real tDCS decreased the number of epileptiform IEDs per min (from 9.46 ± 2.68 after sham tDCS to 5.37 ± 3.38 after real tDCS), p = 0.018, as compared to sham tDCS. GABA was decreased in the EZ after real c-tDCS stimulation as compared to sham tDCS (from 0.129 ± 0.019 to 0.096 ± 0.018, p = 0.02). The reduction in EZ GABA correlated with the reduction in the frequency of epileptiform IED per min (rho: 0.9, p = 0.003). Conclusion: These results provide a window into the antiepileptic mechanisms of action of tDCS, based on local and remote changes in GABA and neural oscillatory patterning responsible for the generation of interictal epileptiform discharges.

Cortical functional topography of high-frequency gamma activity relates to perceptual decision: an Intracranial study.

High-frequency activity (HFA) is believed to subserve a functional role in cognition, but these patterns are often not accessible to scalp EEG recordings. Intracranial studies provide a unique opportunity to link the all-encompassing range of high-frequency patterns with holistic perception. We tested whether the functional topography of HFAs (up to 250Hz) is related to perceptual decision-making. Human intracortical data were recorded (6 subjects; >250channels) during an ambiguous object-recognition task. We found a spatial topography of HFAs reflecting processing anterior dorsal and ventral streams, linked to decision independently of the type of processed object/stimulus category. Three distinct regional fingerprints could be identified, with lower gamma frequency patterns (<45Hz) dominating in the anterior semantic ventral object processing and dorsoventral integrating networks and evolving later, during perceptual decision phases, than early sensory posterior patterns (60-250Hz). This suggests that accurate object recognition/perceptual decision-making is related to distinct spatiotemporal signatures in the low gamma frequency range.

Diabetic brain or retina? Visual psychophysical performance in diabetic patients in relation to GABA levels in occipital cortex.

Visual impairment is one of the most feared complications of Type 2 Diabetes Mellitus. Here, we aimed to investigate the role of occipital cortex γ-aminobutyric acid (GABA) as a predictor of visual performance in type 2 diabetes. 18 type 2 diabetes patients were included in a longitudinal prospective one-year study, as well as 22 healthy age-matched controls. We collected demographic data, HbA1C and used a novel set of visual psychophysical tests addressing color, achromatic luminance and speed discrimination in both groups. Psychophysical tests underwent dimension reduction with principle component analysis into three synthetic variables: speed, achromatic luminance and color discrimination. A MEGA-PRESS magnetic resonance brain spectroscopy sequence was used to measure occipital GABA levels in the type 2 diabetes group. Retinopathy grading and retinal microaneurysms counting were performed in the type 2 diabetes group for single-armed correlations. Speed discrimination thresholds were significantly higher in the type 2 diabetes group in both visits; mean difference (95% confidence interval), [0.86 (0.32-1.40) in the first visit, 0.74 (0.04-1.44) in the second visit]. GABA from the occipital cortex predicted speed and achromatic luminance discrimination thresholds within the same visit (r = 0.54 and 0.52; p = 0.02 and 0.03, respectively) in type 2 diabetes group. GABA from the occipital cortex also predicted speed discrimination thresholds one year later (r = 0.52; p = 0.03) in the type 2 diabetes group. Our results suggest that speed discrimination is impaired in type 2 diabetes and that occipital cortical GABA is a novel predictor of visual psychophysical performance independently from retinopathy grade, metabolic control or disease duration in the early stages of the disease.