A. Participants
Ethical approval was obtained from the St. James’s Hospital / Tallaght University Hospital Joint Research Ethics Committee - Dublin [2014 Chairman’s Action 7]. Experiments were performed in accordance with the Declaration of Helsinki. All participants provided informed written consent to the procedures before undergoing assessment.
Individuals diagnosed with ALS and ALS-frontotemporal dementia (ALS-FTD), according to the revised El Escorial criteria [9] and the Strong criteria [10] were recruited from the Irish National ALS Clinic at Beaumont Hospital in Dublin, Ireland. Those with primary lateral sclerosis, progressive muscular atrophy, flail arm/leg syndromes, other medical conditions, or neurological/neuropsychiatric symptoms were excluded. Age-matched healthy controls (HC), without diagnosed neurological or neuropsychiatric disorders, were recruited from an existing volunteer database [11]. Details of the demographics can be found on Table 1.
Table 1
Demographic profile for controls and patients. The table details the gender proportions, the average ages at recording and, when applicable, disease durations, delays between sessions, site of onset and the number of patients with FTD comorbidity. Numbers show mean ± standard deviation.
Groups | N | Male (%) | Age (years) | Disease duration (months) | ALSFRS-R scores | Site of onset (N) | ALS-FTD diagnosis (N) |
Bulbar | Spinal | Thoracic |
Controls | 78 | 36 | 60 ∓ 12 | / | / | / | / | / | / |
ALS | 99 | 74 | 61 ∓ 11 | 22 ∓ 17 | 37 ∓ 7 | 20 | 72 | 5 | 5 |
B. Experimental setting
High-density (128 channels) resting-state EEG data was recorded from 99 ALS patients and 78 age-matched HC. Data collection was conducted at the Clinical Research Facility in St. James's Hospital, Dublin. EEG recordings took place in a specially designed room, isolated by a Faraday cage to shield against external electric interference. Electrode offsets were maintained within the range of +/- 25mV. Participants were instructed to sit comfortably with their eyes open during the recording session. A large letter X (measuring 6x8 cm2, printed in black on white) served as a focal point for their gaze. EEG signals were captured at a sampling rate of 512 Hz using a 128-channel BioSemi ActiveTwo system from Amsterdam, Netherlands (Honsbeek et al., 1998). A recording was performed in three blocks, each lasting two minutes. The participants' alertness and overall well-being were assessed by the experimenter in brief intervals between each block.
C. EEG pre-processing
The pre-processing procedures followed established protocols outlined in our previous publications [1], [2], [12]. The EyeBallGUI toolbox [13] was applied for visual inspection and quality assessment of the recordings. For the subsequent pre-processing steps, the Fieldtrip Toolbox (version 20190905) [14] was employed.
To identify and discard bad epochs, various criteria were applied, including amplitude, mean shift, variance, and band-variance of spectral power, all assessed against a 3.5 Z-score threshold [15]. Subsequently, the EEG signals were downsampled from 512 to 256 Hz. After resampling, a band-pass filter (one-pass zero-phase FIR: 1-97Hz) and a notch filter (dual-pass third-order Butterworth: 50 Hz, stopband: 1Hz) were applied.
Following baseline correction, a procedure for removing noisy channels was executed, based on both the PREPpipeline and Kohe's work [16], [17]. Channels that were identified as problematic were interpolated using information from adjacent electrodes. Recording sessions where more than 11 channels needed removal were excluded from the study due to their unreliability. On average, controls had 2.6 ± 2.5 channels removed, while patients had 3.6 ± 2.9 channels removed. A common average reference was then applied to the remaining channels.
D. Microstates estimation
The computation of microstates was accomplished using the Microstate EEGlab toolbox [18], as described in our microstate study at sensor-level [3]. In brief, we identified microstates based on quasi-stable topographies of resting-state EEG (1-30Hz). For the combined ALS and HC group, the EEG signals were clustered at peak times of global mean field power (GFP) (where the topographies are the most stable [4]) using modified k-means clustering to obtain four microstate prototypes [18]. This algorithm differs from the original K-means algorithm in that it assigns opposite maps to the same cluster (polarity invariant because an inversion in the direction of the electrical potential on the scalp may occur when the same neural sources produce oscillations within the brain). A Gaussian weighted moving average was added to the GFP signal as a smoothing method (window of five timepoints) [19]. After clustering, each EEG sample was then associated with the microstate prototype it was the most similar to by applying global map dissimilarity [18].
E. Microstates source analysis
In parallel with the microstate processing, the EEG sources were localised using the method described by Dukic et al [1], applied to the full EEG recordings. The source localisation analysis included the use of a linear constraint minimum variance beamformer and an ICBM152 MRI template-based head model [20] for source localization. The source signals were then calculated in 90 brain regions using the automated anatomical labelling atlas [21].
The resulting EEG sources and the sequence of microstates have the same temporal scale, which allows for a direct estimation of the brain areas generating each microstate topography. The resting-state generators underlying EEG microstates were then estimated based on a statistical parametric mapping between the time courses of the microstates and the time courses of the sources [4], [6]. A general linear model (GLM) was applied to fit the 3D brain template for each timepoint to the temporal sequence of microstates [6]:
ηm = Χ ⋅ βm + ε, (1)
with ηm representing a binary function attributing 1 to a time point if it corresponds to the microstate class m, Χ representing the estimated sources (in our 90 brain regions) for each time point and βm representing the weighting factors for all brain regions during a microstate \(\text{m}\). No intercept was included.
A first GLM was applied on the concatenated subjects source map time courses to obtain the brain region weights (or predictors of contribution) βall for each microstate m. Only weights corresponding to significant contributions (FDR correction at 0.05) were kept. The calculation resulted in an estimation of the brain regions contributing the most during each microstate topography in our population.
F. Reproducibility analysis
To evaluate reproducibility, a similar GLM was fitted for each subject \(i\) separately, resulting in the output βi (for each subject i, microstate m). A bootstrapping assessment was, then, performed on the resulting weights, as described by Custo et al. [6]. This bootstrapping assessment involved selecting a random subject, 1000 times (from the pool of existing subjects), and calculated the z-score zs of the estimated contribution maps βi using the mean and standard deviation across the pool of 177 subjects. Next, we computed a reproducibility index by dividing the z-scores of βall by the standard deviation of the 1000 z-score zs.