Electroencephalographic detection of respiratory-related cortical activity in humans: from event-related approaches to continuous connectivity evaluation

Abstract

The presence of a respiratory-related cortical activity during tidal breathing is abnormal and a hallmark of respiratory difficulties, but its detection requires superior discrimination and temporal resolution. The aim of this study was to validate a computational method using EEG covariance (or connectivity) matrices to detect a change in brain activity related to breathing. In 17 healthy subjects, EEG was recorded during resting unloaded breathing (RB), voluntary sniffs, and breathing against an inspiratory threshold load (ITL). EEG were analyzed by the specially developed covariance-based classifier, event-related potentials, and time-frequency (T-F) distributions. Nine subjects repeated the protocol. The classifier could accurately detect ITL and sniffs compared with the reference period of RB. For ITL, EEG-based detection was superior to airflow-based detection (P < 0.05). A coincident improvement in EEG-airflow correlation in ITL compared with RB (P < 0.05) confirmed that EEG detection relates to breathing. Premotor potential incidence was significantly higher before inspiration in sniffs and ITL compared with RB (P < 0.05), but T-F distributions revealed a significant difference between sniffs and RB only (P < 0.05). Intraclass correlation values ranged from poor (−0.2) to excellent (1.0). Thus, as for conventional event-related potential analysis, the covariance-based classifier can accurately predict a change in brain state related to a change in respiratory state, and given its capacity for near “real-time” detection, it is suitable to monitor the respiratory state in respiratory and critically ill patients in the development of a brain-ventilator interface.

breathing is usually under automatic control via rhythmic activity that originates in the medulla (e.g., Feldman and Del Negro 2006), but respiratory motoneurones can also be activated by corticospinal projections (Gandevia and Rothwell 1987; Murphy et al. 1990; Sharshar et al. 2004; Similowski et al. 1996). The cortical areas associated with contractions of the respiratory muscles include the primary motor cortex, premotor cortex, and supplementary motor area (e.g., Colebatch et al. 1991; McKay et al. 2003). A cortical contribution to breathing can be detected by electroencephalographic recordings (EEG) in the form of premotor potentials (or Bereitschaftspotentials; for review, see Colebatch 2007; Shibasaki and Hallett 2006) and were first demonstrated for voluntary sniffs (Macefield and Gandevia 1991). In awake healthy subjects, premotor potentials also precede inspiration in acute and prolonged (1 h) inspiratory loading (Raux et al. 2007b; Tremoureux et al. 2010) and they precede expiration in expiratory loading (Morawiec et al. 2015). EEG evidence of cortical activity during loaded breathing is supported by positron emission tomography (Fink et al. 1996; Isaev et al. 2002) and functional MRI (Raux et al. 2013) studies. Therefore, there is a cortical component to respiratory load compensation. Of note, a reduced motor threshold for diaphragmatic responses evoked by transcranial motor stimulation over the motor cortex in patients with chronic obstructive pulmonary disease is consistent with increased cortical excitability due to chronically increased respiratory load (Hopkinson et al. 2004).

In healthy subjects, the loads that induce cortical activity are externally imposed. However, patients with respiratory disease experience acute intermittent and chronic loading of their respiratory muscles due to, for example, bronchoconstriction or dynamic hyperinflation. In addition to an intrinsic load, critically ill patients requiring mechanical ventilation can experience patient-ventilator asynchrony. Respiratory loads that induce cortical activity can be associated with dyspnea (Morawiec et al. 2015), including during experimental patient-ventilator asynchrony in healthy subjects (Raux et al. 2007a). In mechanically ventilated patients who are able to communicate, dyspnea can be improved with ventilator adjustment (Schmidt et al. 2011), but many critically ill patients cannot verbally report their dyspnea. In this setting, EEG provides a method to detect respiratory-related cortical activity in critically ill patients, which could be used to improve mechanical ventilator settings and to identify and correct dyspnea without the need for direct communication with the patient (see Raux et al. 2007a).

Using event-related approaches to study respiratory-related cortical activity has the major advantage of respiratory specificity, insofar as the EEG signal is processed according to respiratory events. This provides protection against EEG changes concomitant with respiratory changes but not actually due to them. However, analysis of EEG signals by event-related potential methodology has limitations including assumptions of on-going brain activity and the temporal variability of the evoked response as well as high susceptibility to low-frequency movement-related artifacts. This is true for premotor potentials and for alternative methods such as time-frequency analysis (e.g., to reveal event-related desynchronization) that have been applied to EEG signals to demonstrate cortical changes before limb movements (for review, see Makeig et al. 2004; Mouraux and Iannetti 2008; Pfurtscheller and Lopes da Silva 1999). For clinical applications, a major limitation of event-related approaches to detect respiratory-related cortical activity is a large number of trials are required to improve the signal to noise ratio.

The spatiotemporal dynamics of EEG, assessed by interactions of large-scale cortical networks that consider dynamic changes in the brain's intrinsic activity, can be used to investigate brain function on a continuous basis, namely irrespective of discrete events. However, it is not known if these methods are suitable to detect and characterize respiratory-related cortical activity. Recently, we have developed signal processing techniques to discriminate between different breathing conditions using a machine learning based on the EEG covariance matrices (Navarro-Sune et al. 2016). Given its improved temporal resolution, the aim of the present study was to validate the novel covariance-based classifier as a method to identify respiratory-related cortical activity in humans, in other words, to discriminate between resting unloaded breathing and loaded breathing using continuous and rapid EEG processing.

METHODS

The studies were carried out in 17 subjects (9 female) aged 21–29 yr. The subjects gave informed written consent to the procedures which conformed with the Declaration of Helsinki and were approved by the Comité de Protection des Personnes Ile-de-France VI, Pitie-Salpetriere, Paris, France. EEG activity was recorded during three respiratory conditions: resting unloaded breathing, voluntary brisk inhalations (i.e., sniffs), and breathing against an inspiratory threshold load (ITL). Nine subjects (5 female) repeated the protocol for reliability testing. To determine the optimal technique to discriminate between respiratory conditions, EEG were analyzed by premotor potentials, time-frequency analysis of these potentials, and a covariance-based classifier that detects changes in respiratory condition using EEG covariance matrices. The degree of correlation between EEG and airflow was also compared between respiratory conditions, with the hypothesis that a change in correlation would help ascertain the respiratory origin of putative changes in EEG patterns. Within this framework, the degree of correlation between EEG and electrocardiogram was used as control. A subset of these data has been used for statistical and methodological development of the classifier (Navarro-Sune et al. 2016).

Experimental Setup

Subjects were seated in a comfortable chair with neck support and watched a film on a large computer screen during EEG recordings to distract their attention from their breathing and the experimental environment in general. They wore a nose clip and breathed through a mouthpiece connected to a pneumotachograph (Hans Ruldoph), and airflow was measured with a ±2 cmH₂O linear differential pressure transducer (DP45-18; Validyne, Northridge, CA). The pneumotachograph was connected to a two-way valve (2700 series; Hans Ruldolph). End-tidal CO₂ was measured at the expiratory port by an infrared CO₂ gas analyzer (Servomex, Plaine Saint Denis, France), and when required, an ITL (Respironics) was attached to the inspiratory port of the two-way valve. Opening airway pressure was measured close to the mouth with a ±225 cmH₂O differential pressure transducer (DP15-34; Validyne). An electrocardiogram (ECG) was measured with two surface electrodes placed on the left side of the chest, above and below the nipple. Respiratory and ECG signals were sampled at 2 kHz.

An active electrode system (ActiCap; BrainProducts) was used to record scalp EEG from 32 electrodes positioned according to the international EEG 10–20 system, including 2 electrodes on the ear lobes. Electrooculographic activity (EOG) was monitored with an electrode positioned under the right eye. The recordings were amplified and digitized at 2.5 kHz (BrainAmp; BrainProducts). Online, recordings were referenced to FCz with a ground at FPz. EEG signals were time locked to respiratory signals with simultaneous digital trigger pulses that were automatically generated from a threshold crossing of inspiratory airflow and manually verified offline by visual inspection.

Experimental Protocol

Subjects performed three conditions (duration: 10–20 min each): 1) resting unloaded breathing, in which no instructions were given to the subject in relation to their breathing; 2) self-paced brisk voluntary inhalations, akin to “sniffs” but made through the mouth due to the experimental apparatus (subjects were instructed to perform the “sniffs” at the start of the usual inspiratory phase of the respiratory cycle, but not every cycle); and 3) breathing with an ITL set to ∼23 cmH₂O (range: 18–25 cmH₂O), placed on the inspiratory port of the two-way valve. Subjects always performed the resting breathing condition first to minimize awareness that their breathing was being manipulated. The sniff and ITL conditions were then performed in random order. At the end of each condition, subjects rated their level of respiratory discomfort during the condition using a visual-analog scale that ranged from zero: “no breathlessness or discomfort” to 10: “maximal breathlessness or discomfort.” There was a break of 5–10 min between conditions.

Reliability.

Nine subjects repeated the same protocol of resting unloaded breathing, voluntary sniffs, and ITL, quasirandomized as in the first visit. The interval between visits averaged 18.1 [7.3; 95% confidence interval (CI)] days with a range of 6–38 days. The ITL was set to the same pressure as the first visit.

Data Analysis

Respiratory variables.

Tidal volume, inspiratory time, peak negative mouth pressure, end-tidal CO₂, respiratory rate, and ventilation were measured for resting breathing and ITL for recordings from nine subjects on the 2 days of the protocol.

Premotor potential analysis.

Offline, recordings were resampled at 250 Hz, referenced to linked-ear electrodes, and band-pass filtered (0.05–10 Hz). Before ensemble averaging, the EEG was divided into 3.5-s epochs, 2.5 s before and 1 s after the onset of inspiration, based on the onset of inspiratory flow (resting breathing and sniffs) or negative mouth pressure (ITL). Individual segments were visually inspected and rejected from the average if they exhibited artifact, e.g., large deviations from baseline or intense EOG activity. One subject had such frequent and intense EOG activity such that after removal of the artifact-affected segments, 25 or less segments remained for the ensemble averages. Therefore, data from this subject were removed from the premotor potential analysis. For the remaining 16 subjects, 72 ± 12 segments (means ± SD, range: 38–80) were averaged for each condition.

Average EEG traces were examined for the presence of an inspiratory premotor potential as described previously (Raux et al. 2007b), i.e., a slow increasingly negative shift in the EEG signal starting from ∼1 s before inspiration. If present, the preinspiratory latency of each premotor potential was measured as the duration (in ms) between the onset of EEG negativity and the start of inspiration and the amplitude was measured (in μV) at this “zero flow” point.

Premotor potential analysis was performed on Cz and FCz derivations. Although inspiratory premotor potentials have previously been reported for Cz only (e.g., Macefield and Gandevia 1991; Raux et al. 2007b; Tremoureux et al. 2010), the observed potentials were more frequent and were larger in amplitude at FCz. Therefore, FCz was used as the most likely to reflect activity in the premotor and supplementary motor areas.

Time-frequency analysis.

Time-frequency analysis was performed for FCz on same EEG epochs as premotor potential analysis (n = 16 subjects), but signals were resampled at 1,000 Hz. Wavelet time-frequency distributions were obtained using the wavelet transform (Tallon-Baudry et al. 1997). The EEG signal x(t) at FCz was convolved with a complex Morlet's wavelet function defined as:

$$w(t,\hspace{0.17em}{f}_0)=A\hspace{0.17em}\exp \hspace{0.17em}\left(\frac{-t^2}{2{\sigma }_t^2}\right)\times \exp \hspace{0.17em}(i2\pi f_{0}t)$$

Wavelets were normalized and thus:

$$A={({\sigma }_t\sqrt{\pi })}^{-1/2}$$

The width of each wavelet function m = f₀/σ_f was chosen to be 7, where σ_f = 1/2π σ_t. The parameter m corresponds to the number of significant cycles of the wavelet function and defines the trade-off between temporal and frequency resolution (i.e., for larger values of m, the frequency resolution increases but the temporal resolution decreases). A m value larger than 5 ensures the necessary admissibility condition for the wavelet function (an oscillating function is admissible if it has finite energy and zero mean value). Time-frequency contents were represented as the energy of the convolved signal:

$$E(t,f_0)={\left|w(t,\hspace{0.17em}{f}_0)\otimes x(t)\right|}^2$$

To detect preinspiratory changes in temporal and spectral power of the EEG, a “baseline” period of a 300-ms duration was defined. For resting breathing and ITL, the baseline period commenced at peak expiratory flow of the preceding breaths, based on the average airflow signal of the analyzed segments. For the sniff condition, as the maneuvers were self-paced and interspersed with spontaneous breaths, the expiratory time was shortened and likely more variable. For the sniff condition, baseline was considered to be −1,900 to −1,600 ms before the onset of inspiratory flow, estimated from the average latency +99% CI of sniff premotor potentials (Macefield and Gandevia 1991).

Following preprocessing (each trial was assigned zero mean, unitary variance, and corrected by the baseline), single-trial time-frequency maps were computed. For each frequency, single-trial time-frequency maps were corrected by the mean and standard deviation power during the baseline period. P(t,f) was then averaged across all trials for each condition.

To quantify the time-frequency maps in the different conditions, a region of interest (ROI) between 0 and 2 Hz and −1,000 and 0 ms was selected. Visually, this low-frequency band showed the greatest differentiation between conditions and the time period was common to all conditions as the mean preinspiratory latency of the premotor potentials was ∼1,000 ms or less (see Fig. 3). The mean value of the top 20% of time-frequency values in this ROI was determined for each subject and condition (Mouraux and Iannetti 2008). The same ROI was used for reliability measures.

Fig. 3.

Group data for premotor potential and time-frequency analyses. A: the incidence and mean [95% confidence interval (CI)] latency, amplitude of inspiratory premotor potentials at FCz in RB, voluntary sniffs, and breathing with an ITL for the group of subjects (n = 16). For mean preinspiratory latency and amplitude, the number of subjects for mean data in each condition is indicated in the bars. B: for all 16 subjects and the same epochs of EEG, the average top 20% pixel values from time-frequency maps are shown at right. Median interquartile range ([IQR]) is shown for pixel value. *Statistically different to RB.

Covariance-based classifier.

EEG data were also analyzed using an algorithm that classifies brain activity in different conditions using a semisupervised approach (Navarro-Sune et al. 2016). We tested for altered brain activity in “modified” breathing conditions (i.e., ITL and sniffs), compared with reference activity (e.g., resting breathing) and detection was performed using covariance matrices, a very appropriate feature to capture the spatiotemporal structure of the EEG (Barachant et al. 2012; Varela et al. 2001).

EEG signals from frontal and central channels (F3, F4, C3, C4, FCz, Fz, Cz, FC2, and FC6) were downsampled to 250 Hz and band-pass filtered (8–24 Hz) to enhance motor cortical activity (or mu rhythm) found in this frequency band (Pfurtscheller and Lopes da Silva 1999). During the learning process, the reference period (20% of resting breathing) was divided into a series of signals X_m (m = 1, ․ ․ ․, M number of segments). Let X = X_m $\in$ R^{P × N} a reference period of EEG signals recorded during N temporal samples of a time window equal to 5 s with 50% overlap and P EEG channels (see Fig. 1). Each X matrix was centered and normalized. Then, spatial covariance matrices C were estimated for each X by: C = (X X^T)/trace (X X^T).

Fig. 1.

Schema to demonstrate detection of an altered “brain state” with the covariance-based classifier. A: covariance matrices are computed from 1, ․ ․ ․ P EEG signals and 1, ․ ․ ․ N 5-s sliding windows to capture the spatiotemporal dynamics of the EEG. This is first done for a reference period of resting breathing to generate prototypes of this “brain state.” Covariance matrices are then computed from recordings during different breathing conditions: unloaded resting breathing (RB), breathing with an inspiratory threshold load (ITL), or voluntary sniffs. The Riemannian distance (dist_R) to the prototype for RB and ITL is depicted in A, bottom (see also histogram in Fig. 4A). B: a 3-dimensional visualization of covariance matrices (C¹, ․ ․ ․ C^M) as points on an imaginary manifold. The distance between matrices, for example, for RB (shown in red) and ITL (shown in blue), are then compared with a prototype (S). If the distance is greater than a threshold ($\mathrm{T}=\overline{MAD\hspace{0.17em}\text{dist}\hspace{0.17em}(S,\hspace{0.17em}{C}^1,\hspace{0.17em}\dots \hspace{0.17em}{C}^M)}$ ; where MAD is the median absolute deviation; see methods) then there is an altered brain state. The ability of the covariance-based classifier to detect a change in brain state is tested using the receiver operator characteristic curve (ROC; see methods for details). The inset highlights the difference between Riemannian (used in the classifier) and Euclidean geometry to compute the distance between 2 points.

Since C are symmetric positive definite matrices that lie in a Riemannian manifold, the distance between two matrices C¹ and C² can be defined as:

$${\text{dist}}_{\mathrm{R}}({\text{C}}^1,\hspace{0.17em}{\text{C}}^2)={\left[{\sum }_{k=1}^p{\log }^2{\displaystyle {\propto }_k}\right]}^{1/2}$$

α_k the k-th eigenvalue of [(C¹)⁻¹ (C²)] k = 1:p.

To represent the reference period, prototypes S_j, j = 1, ․ ․ ․, L, were determined from the distribution of the covariance matrices Cⁱ, i = 1: M. Each prototype represented a subclass of covariance matrices and was estimated by the Karcher average of covariance matrices. The calculation procedure of prototypes was based on a modified K means clustering algorithm, where the geometric mean of M matrices was obtained by the following formula for calculating the average Karcher M_R (Moakher 2005):

$${\text{M}}_{\text{R}}\hspace{0.17em}({\text{C}}^1,\hspace{0.17em}{\text{C}}^2,\hspace{0.17em}․\; ․\; ․,\hspace{0.17em}{\text{C}}^{\text{M}})=\arg \hspace{0.17em}\min \hspace{0.17em}{\displaystyle {\sum }_{\text{m}=1}^M{\text{dist}}_{\text{R}}\hspace{0.17em}({\text{C}}^{\text{m}},\hspace{0.17em}\text{C})}$$

Calculating M_R may be performed using a gradient descent procedure that converges rapidly (Pennec et al. 2006).

During the testing process, the classifier compares the covariance matrices from both resting and altered breathing with the prototypes learned from the reference period.

First, the distribution of the distance from all the covariance matrices C ⁱ, i = 1: M, to the closest prototype S was calculated, and then the absolute values of deviations from the median of the distribution were determined. The average deviation was used to estimate the rejection threshold beyond which the covariance matrices no longer belonged to the reference situation, i.e., dispersion from prototypes or altered brain activity.

The application of a threshold, based on the dispersion of the matrices corresponding to the reference period, allowed detection of a characteristic change in EEG activity in a situation of modified breathing.

For comparison to detection by EEG, a similar procedure was performed using the airflow signal only. Here, an amplitude feature of the airflow signal for 5-s windows with 50% overlap was calculated. The distance between the time-windows in the reference period and modified breathing was used to classify breathing states (for details, see Navarro-Sune et al. 2016).

We tested the sensitivity of the different detection methods by a 10-fold cross validation. The reference period of resting breathing was divided into 10 equal parts. Comparison between nine of these parts from the reference period and the test set from the modified situations of ITL and sniffs was then performed. This procedure was repeated nine times to take account all combinations of nine parts from the reference period. This allowed us to create receiver operating characteristic curves (ROCs) and a complete sensitivity/specificity report. The ROC curve is a fundamental tool for diagnostic test evaluation and the area under the curve (AUC) quantifies the sensitivity/specificity ratio. A test with perfect discrimination has an AUC equal to 1 and a random discrimination has a value equal to 0.5.

EEG correlations.

We tested for correlation between the power of EEG signals at FCz, Cz, Fz, C3, and C4 with airflow. As for the covariance-based classifier, 5-s windows with 50% overlap were used. Autocorrelation and time trends were present in our time series, indicating that low-frequency variability was important and would increase the chance that spurious correlations would occur in standard inference tests. Therefore, after de-trending, we used a procedure to examine correlations using the method recommended by Pyper and Peterman (1998), which adjusts degrees of freedom to account for autocorrelation and computes the Fisher-transformed correlation (corrected for autocorrelation) that has a Gaussian distribution with mean 0 and unitary variance. As a control, correlations between EEG and ECG were computed in the same way.

Statistics

Group data are presented as mean (95% CI) or median interquartile range ([IQR]) for parametric and nonparametric data, respectively. The H statistic [degrees of freedom (df)] is shown for unpaired data and the F statistic or χ²-value (df) is shown for parametric and nonparametric repeated measures tests, respectively. The criterion for statistical significance was taken as P < 0.05.

A Kruskal-Wallis one-way ANOVA on ranks with pairwise post hoc testing using the Tukey's test was performed to compare the incidence, amplitude, and latency of inspiratory premotor potentials between conditions.

All other data were repeated measures. Therefore, for the group of 17 subjects, conditions (3 levels) were compared one-way repeated measures (RM) ANOVA for parametric data or with Friedman RM ANOVA on ranks for nonparametric data and pairwise post hoc testing with Tukey's test. For the nine subjects who repeated the protocol, two-way RM ANOVAs were used to compare between condition (2 or 3 levels) and day (2 levels) with Tukey's test for post hoc pairwise comparisons. Latency, amplitude, and top 20% data failed normality testing (Shapiro-Wilk, P < 0.05) and thus data were transformed with the rank procedure before statistical testing by two-way RM ANOVA. Two-way RM ANOVA was also used to compare between EEG- or airflow-based detection of brain state with condition (ITL and sniff) and type (flow or EEG) as factors.

The correlation between EEG and airflow during resting breathing and ITL are nonoverlapping correlations from the same sample population. Therefore, they were compared using the Pearson-Filon test for Fisher-transformed correlations (Raghunathan et al. 1996). The same procedure was used for EEG-ECG correlation.

For the nine subjects who repeated the protocol, the reliability of respiratory variables and EEG analyses was assessed using intraclass correlation (ICC) analysis (mixed model, single measures). The reliability was assessed independently for each condition. t-Tests were used to compare inspiratory time and tidal volume when an inspiratory premotor potential did or did not occur during resting breathing in these nine subjects.

RESULTS

Premotor Potentials

Figure 2A shows typical examples of the averaged preinspiratory FCz EEG signal in the different respiratory conditions. As expected, inspiratory premotor potentials were present in the sniff and ITL conditions, but ensemble averaged EEG was unchanged during resting unloaded breathing. For the group of 16 subjects, all had an inspiratory premotor potential during sniffs, 13 during ITL, and 3 subjects during resting breathing such that premotor potential incidence significantly varied with condition [H₍₂₎ = 25.5, P < 0.05; Fig. 3]. Premotor potentials were more common in sniff and ITL conditions compared with resting breathing (P < 0.05), but there was no difference between sniffs and ITL (P > 0.05). The latency [F_(2,29) = 2.3, P = 0.1] and amplitude [F_(2,29) = 1.8, P = 0.2] of premotor potentials did not differ with condition (Fig. 3).

Fig. 2.

Representative EEG data during RB, voluntary sniffs, and breathing with an ITL. A: epochs of EEG at FCz (linked earlobe reference; A1A2), time locked to the onset of inspiratory airflow or negative airway pressure, were averaged to investigate the presence of inspiratory premotor potentials. For this typical subject, the potential was absent in RB, but present in the sniff and ITL condition as evidenced by the increasing negativity of the EEG signal ∼1,000 ms before the onset of inspiration. B: for the same subject and same epochs of EEG as the average data in A (numbers for both analysis types given above each map), time-frequency maps were computed. There were differences in the power of the low-frequency components. To quantify these differences, the average value of the top 20% of pixels from a region of interest (see methods) was calculated. The color scale is −5 to 20 for all panels.

Time-Frequency Maps

Computation of the premotor-related temporal and frequency changes in EEG revealed differences between respiratory conditions. Differences were evident in a ROI of low frequencies (0–2 Hz) at a similar latency to premotor potentials, i.e., from ∼1 s before the onset of inspiration. Typical data from one subject are shown in Fig. 2 and quantification of group data in Fig. 3. Consistent with the premotor potential incidence, there were significant differences between respiratory conditions, quantified as the mean top 20% pixel value in the ROI [χ²₍₂₎ = 11.4, P < 0.01; see methods]. Compared with resting breathing, the average pixel value was significantly greater in voluntary sniff (P < 0.05) but not in ITL. Thus the power of the EEG in a low-frequency ROI was greatest in the voluntary sniff condition and smallest in resting breathing.

Covariance-Based Classifier

Detection of altered breathing condition by the classifier was assessed by the area under the ROC curve. With resting unloaded breathing as the reference period, the AUC averaged 0.85 (0.05) for ITL and 0.83 (0.07) for the sniff detection, respectively. Classification based on the airflow signal only averaged 0.57 (0.15) for ITL but was 0.83 (0.09) for the sniffs. There was a significant effect of detection type [i.e., EEG or airflow based; F_(1,16) = 9.1; P < 0.05], condition [ITL or sniff; F_(1,16) = 6.4; P < 0.05], and an interaction [F_(1,16) = 9.5; P < 0.05]. Post hoc testing revealed that flow-based detection was better in the sniff condition than ITL (P < 0.05) and that EEG-based detection was superior to airflow-based detection for ITL only (P < 0.05; Fig. 4).

Fig. 4.

Detection of altered breathing condition by covariance-based classifier on EEG and airflow. A: representative histogram of the Riemannian distance between EEG covariance matrices in RB and ITL. As assessed by the area under the ROC (see methods), this subject had the best detection (0.99). Not all subjects had perfect discrimination as shown by the mean data in B. However, for individual subjects, the poorest detection was 0.65 using EEG compared with 0.18 when detection was based on airflow. B: mean (95% CI) area under the ROC for detection between a reference period of RB and ITL (left) or RB and sniffs (right). Airflow- and EEG- based detection are shown in white and black bars, respectively. The EEG method had superior detection between RB and ITL. *Significantly different to airflow-based detection.

EEG Correlations

To confirm that changes in brain activity during ITL detected by the EEG-based classifier was related to breathing, correlations between EEG and airflow were assessed. All five central EEG channels had high EEG-airflow correlations, ranging between 0.61 and 1.0 in resting breathing and 0.73 and 1.19 in ITL. Maximal mean correlations in both conditions were for FCz and thus statistical testing was performed on this channel only. For the group of 17 subjects, mean EEG-airflow correlation was 0.99 (0.07) in ITL, significantly higher than 0.83 (0.07) in resting breathing (z value: −8.44, P < 0.05). As a control, we tested the correlation between EEG and ECG at FCz and it did not differ between breathing conditions (z value: 0.34, P = 0.7). Mean EEG-ECG correlation was 0.89 (0.13) during resting breathing and 0.87 (0.10) in ITL.

Respiratory Discomfort

On a 10-cm scale with 10.0 cm as maximal, ratings of respiratory discomfort differed with condition [F_(2,32) = 59.0; P < 0.05]. Respiratory discomfort was greatest during ITL and averaged 4.7 (0.9) cm. This was significantly higher than 0.6 (0.4) cm during resting breathing and 1.5 (0.8) cm for sniffs (P < 0.05 for both).

Reliability

Data for the subgroup of subjects who repeated the same protocol of resting breathing, sniffs, and ITL on 2 days are shown in Tables 1 and 2. The same between-condition differences were observed for premotor potential and time-frequency analyses, and there was no difference in any measure between days. The interclass correlation values for the reliability of these measures (where applicable, i.e., adequate data) ranged from poor (−0.14) to excellent (1.0).

Table 1.

Reliability of premotor potential and time-frequency analysis of respiratory-related cortical activity

	Resting Breathing			Sniffs			ITL			F Statistic and P Value
	Day 1	Day 2	ICC	Day 1	Day 2	ICC	Day 1	Day 2	ICC	Condition	Day
Incidence	3/9	3/9	—	9/9*	9/9*	1.0	7/9 *	9/9 *	—	F(2,48) = 19.8; P < 0.05	F(1,48) = 0.6; P = 0.4
Amplitude, μV	3.8 (3.0)	4.3 (3.5)	—	5.3 (2.1)	7.9 (3.2)	0.65	4.8 (2.9)	6.5 (5.3)	0.56	F(2,34) = 2.5; P = 0.1	F(1,35) = 1.0; P = 0.3
Latency, ms	1083.3 (851.3)	825.0 (887.1)	—	878.8 (439.5)	836.7 (357.3)	0.28	602.9 (277.2)	898.9 (411.7)	0.52	F(2,35) = 0.6; P = 0.6	F(1,35) = 0.07; P = 0.8
Top 20% pixel	10.9 (4.3)	9.5 [8.5–14.0]	0.27	32.8* (17.3)	42.7* [26.0–128.2]	−0.14	21.9 (13.3)	24.1 [7.2–42.5]	0.40	F(2,16) = 10.8; P < 0.05	F(1,8) = 1.9; P = 0.2

Incidence and mean [95% confidence interval (CI)] or median (interquartile range [IQR]) data for the subgroup of subjects who repeated the protocol of resting breathing, voluntary sniffs, and breathing with an inspiratory threshold load (ITL) on 2 days. Averaged EEG activity was compared by premotor potential incidence, latency and amplitude or quantified by the mean top 20% pixel value in a common region of interest (see methods) in time-frequency maps. Intraclass correlation (ICC) values as well as the F statistic and P values from two-way repeated measures (RM) ANOVA are shown.

^*Statistically different from resting breathing in post hoc tests (P < 0.05).

Table 2.

Reliability of EEG correlations and respiratory variables in resting breathing and ITL

	Day 1	Day 2	ICC	Day 1	Day 2	ICC	Day 1	Day 2
	Resting breathing			Inspiratory threshold load			Z value and P value
EEG correlations
FCz-airflow	0.79 (0.11)	0.79 (0.09)	0.79	0.98 (0.09)*	0.95 (0.11)*	0.45	−5.13; P < 0.05	−2.18; P < 0.05
FCz-ECG	0.91 (0.16)	0.91 (0.08)	0.44	0.88 (0.11)	0.95 (0.12)	0.21	0.86; P = 0.4	−0.48; P = 0.6
	Resting breathing			Inspiratory threshold load			F statistic and P value: condition	F statistic and P value: day
Respiratory variables
Inspiratory time, s	1.84 (0.28)	1.74 (0.38)	0.89	4.66 (1.46)	4.41 (2.47)	0.53	F(1,7) = 15.2; P < 0.05	F(1,7) = 0.7; P = 0.4
Tidal volume, liter	0.53 (0.10)	0.53 (0.16)	0.89	1.04 (0.36)	0.93 (0.27)	0.59	F(1,7) = 18.4; P < 0.05	F(1,7) = 0.9; P = 0.4
Pressure, cmH2O	−1.44 (0.24)	−1.43 (0.27)	0.34	−22.36 (4.12)	−23.37 (4.66)	0.87	F(1,7) = 131.8; P < 0.05	F(1,7) = 1.0; P = 0.3
End-tidal CO2†	5.52 (0.70)	5.41 (0.25)	0.82	4.33 (0.53)	4.55 (0.63)	0.78	F(1,7) = 22.1; P < 0.05	F(1,7) = 0.6; P = 0.5
Respiratory rate, cycles/min	14.49 (2.43)	15.43 (3.33)	0.90	9.10 (2.67)	9.35 (4.32)	0.59	F(1,7) = 15.8; P < 0.05	F(1,7) = 1.4; P = 0.3
Ventilation, l/min	7.20 (0.78)	7.68 (1.36)	0.76	8.80 (3.62)	7.99 (3.06)	0.61	F(1,7) = 0.6; P = 0.5	F(1,7) = 0.02; P = 0.9

Mean (95% CI) correlation between FCz and airflow was significantly greater during ITL than resting breathing whereas the correlation between FCz and ECG was similar. Respiratory variables are shown for 8 of the 9 subjects who repeated the protocol. One subject had missing data on day 2 due to a technical error. ICC values and statistics are shown.

^*Statistical different to resting breathing in post hoc tests (P < 0.05); †statistical interaction between condition and day [F_(1,7) = 6.0; P < 0.05].

Changes in brain activity by covariance-based classifier averaged 0.90 (0.05) and 0.85 (0.09) for ITL and sniff conditions, respectively, on day 1 and 0.80 (0.07) and 0.84 (0.12), respectively, on day 2. There was no main effect of condition [F_(1,8) < 0.001, P = 1.0] or day [F_(1,8) = 1.7, P = 0.2]. A statistically significant interaction between condition and day [F_(1,8) = 11.6, P = 0.009] revealed superior detection of ITL on day 1 compared with day 2 (P = 0.05). This may be due to subjects not being naïve to manipulation of their breathing during the resting breathing condition (that is used as a reference) on day 2. Despite good detection, as indicated by the high ROC values, reliability assessment showed poor ICC values of −0.2 and 0.5 for ITL and sniff detection, respectively.

Mean and ICC values for EEG-airflow and EEG-ECG correlations for resting breathing and ITL and respiratory variables from the corresponding recording sessions are shown in Table 2. As expected, ITL changed respiratory variables, except for ventilation, which was similar to that in resting breathing. There was no effect of day on any variable and the ICC values ranged from 0.34 to 0.90. For the occurrences when an inspiratory premotor potential was observed during resting breathing, inspiratory time was ∼40% longer and tidal volume ∼40% larger compared with these parameters when a premotor potential was not observed (P < 0.05 for both).

DISCUSSION

The most significant outcome from this study is that a change in breathing condition can be detected accurately from the spatiotemporal dynamics of EEG signals. Compared with event-based approaches, this method considerably reduces the number of ventilatory cycles needed to detect a change and therefore improves processing speed. In healthy subjects, detection of loaded breathing was superior using EEG signals rather than airflow. A higher EEG-airflow correlation in loaded compared with resting breathing confirms the change in brain state detected by the covariance-based classifier is related to respiration. Although premotor potential incidence could also discriminate between breathing conditions for the group of subjects, discrimination was not possible in subjects who exhibited a premotor potential during resting breathing. The long preprocessing time of this averaging technique is a limitation for clinical applications. Accurate and rapid detection of loaded breathing in healthy subjects, a model of lung disease, respiratory failure, and ventilator fighting are important steps towards the development of a brain-ventilator interface.

Loaded Breathing, Altered Brain State, and Neurophysiological Substrates

The Bereitschaftspotential (or premotor potential) is a slow negativity that reflects synchronized changes in the postsynaptic potentials of cortical neurones (Colebatch 2007; Shibasaki and Hallett 2006). Respiratory-related cortical activity has been demonstrated previously using this time-locked EEG averaging technique (see Introduction). As for limb movements (for review, see Shibasaki and Hallett 2006) premotor potentials are observed during sniffs and inspiratory loaded breathing in EEG deviations considered to reflect activity in the supplementary motor area and premotor and primary motor cortices. A recent study has demonstrated similar current sources (or dipoles) for premotor potentials during voluntary sniffs and finger flexions using principal component analysis (Jeran et al. 2013). Cortical activity during loaded breathing is proposed to reflect cortical automatization rather than breath-to-breath modulated “voluntary” inspirations as would occur in the sniff condition. This notion derives from the comparison of functional MRI patterns observed during single-breath inspiratory loading and continuous inspiratory loading, which showed a reduction in overall brain activity but reinforcement of supplementary motor area activity during continuous loading (Raux et al. 2013). This is similar to what is observed during “overlearning” or cortical automatization (Lehericy et al. 2005; Wu et al. 2004, 2008).

We had anticipated that time-frequency decomposition would reveal differences in alpha and beta activity between respiratory conditions, consistent with event-related desynchronization as described for voluntary limb movements (for review, see Pfurtscheller and Lopes da Silva 1999). Breathing is rhythmic and the period between trials is relatively fixed. Therefore, unlike limb movement studies, we were unable to have a baseline that was removed in time from our stimulus or a long interevent interval of at least 10 s as recommended (Mouraux and Iannetti 2008; Pfurtscheller and Lopes da Silva 1999). Rather, we observed and quantified differences in low-frequency EEG power (see Figs. 1 and 2) that likely reflects premotor potentials in the different conditions. As time-frequency analysis also requires multiple trials to be averaged and the temporal resolution of low frequencies can be uncertain (Mouraux and Iannetti 2008), it provides no greater advantage than premotor potentials to discriminate between loaded and unloaded breaths in a clinical setting.

Rather than test for time-locked responses at individual EEG electrodes, the covariance-based classifier tests for changes in brain activity as a result of altered interactions in large-scale cortical networks. Functional interactions between brain regions are tested by comparing local neural synchronization or variance (i.e., within an EEG derivation) in the diagonal terms (or power) of spatial covariance matrices and longer distance characteristics of the synchronization of the neural network (i.e., between EEG derivations) by the off-diagonal terms (see Supplementary Video; Supplemental Material for this article is available online at the Journal website). Here, matrices were computed from central and frontal EEG channels with similar detection accuracy for both sniffs and loaded breathing suggesting the networks activated in these “brain states” are similar and include the supplementary, prefrontal, and primary somatosensory motor areas. With the use of the covariance-based classifier, differences in the electrodes that give the best detection of loaded breathing or sniffs suggest the exact pattern of cortical activity also differs slightly between these respiratory tasks (Navarro-Sune et al. 2016; see also Isaev et al. 2002). There may be distinct, but overlapping, networks for these breathing conditions as described for breathing and speech (Loucks et al. 2007).

As the covariance-based classifier used segments of EEG that were not synchronized with respiration, the altered brain state may not necessarily reflect central network activity corresponding to a change in breathing condition. Respiratory muscles are activated in voluntary and loaded breaths, (see Hudson et al. 2011; Raux et al. 2007b) and the final motor output is manifested as a change in airflow. A coincident increase in the correlation between EEG and airflow in loaded breathing compared with resting breathing suggests the altered brain state as detected by the covariance-based classifier was due to breathing-related activity.

Methodological Considerations (Including Reliability)

Premotor potentials were observed during resting breathing in 3 of 17 subjects, at a similar incidence as described previously for healthy subjects (see Tremoureux et al. 2014) and most likely due to an awareness or volitional control of breathing due to the respiratory apparatus despite modest efforts to prevent this. An increase in inspiratory time and tidal volume during recordings when a premotor potential was observed during resting breathing is consistent with previous reports of cortical activation during augmented breaths (Jutand et al. 2012). Although task-related differences in the properties of premotor potentials have been reported in limb movements, comparable amplitudes and latencies of premotor potentials in different respiratory conditions are consistent with previous studies of respiratory-related cortical activity (Launois et al. 2015; Morawiec et al. 2015; Raux et al. 2007b; Tremoureux et al. 2010).

For premotor potentials and time-frequency analysis, only EEG epochs free of intense EOG activity and large deviations due to movement were included. For the covariance-based classifier, a band-pass filter of 8–24 Hz would have removed major components of ocular and movement artifact. Furthermore, frontal polar electrodes were excluded and any vertical ocular artifact is likely to have been reduced by 80% at the frontal electrodes used in the covariance-based classifier (Croft and Barry 2000).

For nine subjects, reliability was poor for premotor potential and time-frequency analysis measures, except for premotor potential incidence during the sniff condition. These data are consistent with reproducibility of preinspiratory potentials from three subjects during voluntary sniffs (Raux et al. 2007b) and poor reliability for premotor potential latency in six subjects who repeated a protocol of finger extension (Matsuhashi and Hallett 2008). A limitation of premotor potentials is that cortical activity cannot be quantified in all subjects and all conditions if the potential is absent. However, quantification of premotor potentials via low-frequency power in time-frequency maps did not provide a more reliable method to detect changes in different respiratory conditions. The covariance-based classifier directly compares and tests for differences between conditions with good accuracy (AUC >0.8). Furthermore, it can detect a change in all subjects, in contrast to event-related potentials that cannot discriminate between unloaded and loaded breathing when subjects evidence a premotor potential in resting breathing.

Development of a Brain-Ventilator Interface

Premotor potential and time-frequency analyses used in this study require a large number of trials to enhance the signal-to-noise ratio. This is adequate for physiological explorations in the laboratory when data can be analyzed offline. However, for clinical applications such as a brain-ventilator interface that uses EEG to detect a change in breathing, an improved detection method is required. Here, in healthy subjects, we could accurately detect a change between resting and loaded breathing using a semisupervised algorithm with short preprocessing time (see also Navarro-Sune et al. 2016). Although the current analysis was performed offline, pilot testing in our laboratory has confirmed online detection is possible within a quasi-real-time domain (>0.01 s for 5-s windows) following classifier set-up (i.e., learning the reference period) of ∼1 min (Navarro-Sune et al. 2016). Thus, compared with premotor potential analysis, detection of cortical activity is close to “real-time,” which makes the covariance-based classifier suitable for monitoring of respiratory state. Rapid detection is particularly important in critically ill patients in whom respiratory state can be labile due to changes in respiratory mechanics or neural activity (e.g., due to changes in sedation or pain) that alter the brain-lung interface.

Nine frontal and central EEG channels were used to detect a change in breathing condition. Future studies should investigate how to optimize intersubject detection with a minimal number of electrodes, for example, by controlling for potential nonrespiratory variations in the EEG such as head movement artifact (see O'Regan and Marnane 2013). The proposal of a brain ventilator interface that uses EEG to monitor medullary respiratory activity to trigger a mechanical ventilation (Grave de Peralta et al. 2013) requires further technical development. The central neural drive can already be used to drive mechanical ventilation via diaphragm EMG in neurally adjusted ventilator assistance (e.g., Schmidt et al. 2015) and furthermore, EEG monitoring of medullary activity does not consider neural drive to the respiratory muscles that originates in cortical circuits. Respiratory-related cortical activity has been shown to be associated with respiratory discomfort in cases of extrinsic respiratory loading (Morawiec et al. 2015; Raux et al. 2007a), and this association is to be confirmed for intrinsic loads experienced by respiratory and critically ill patients. Recently, we proposed a brain-ventilator interface that uses EEG to detect loaded breathing and respiratory discomfort (Navarro-Sune et al. 2016). Here, we demonstrate for the first time that respiratory-related cortical activity associated with loaded breathing in healthy subjects can be detected accurately and with improved time resolution.

GRANTS

This work was supported by the program Investissement d'Avenir EMMA-0030 and ANR-10-AIHU 06 of the French Government and by the grant Legs Poix from the Chancellerie de l'Universite de Paris, France. X. Navarro-Sune's postdoctoral position within UMRS1158 is funded by Air Liquide Medical Systems. A. L. Hudson was supported by an National Health and Medical Research Council of Australia (Australia) Early Career Fellowship.

DISCLOSURES

X. Navarro-Sune is employed by Air Liquide Medical Systems S.A., France. T. Similowski, M. Raux, M. Chavez, J. Martinerie, and P. Pouget are listed as inventors for patents using EEG to adjust mechanical ventilation. UMRS1158 and ALMS have a contract to develop a brain ventilator interface.

AUTHOR CONTRIBUTIONS

Author contributions: A.L.H., P.P., M.R., M.C., and T.S. conception and design of research; A.L.H. performed experiments; A.L.H., J.M., and M.C. analyzed data; A.L.H., X.N.-S., J.M., P.P., M.R., M.C., and T.S. interpreted results of experiments; A.L.H. and X.N.-S. prepared figures; A.L.H. drafted manuscript; A.L.H., X.N.-S., J.M., P.P., M.R., M.C., and T.S. edited and revised manuscript; A.L.H., X.N.-S., J.M., P.P., M.R., M.C., and T.S. approved final version of manuscript.