Unsupervised detection and removal of muscle artifacts from scalp EEG recordings using canonical correlation analysis, wavelets and random forests

OBJECTIVE

This paper proposes supervised and unsupervised algorithms for automatic muscle artifact detection and removal from long-term EEG recordings, which combine canonical correlation analysis (CCA) and wavelets with random forests (RF).

METHODS

The proposed algorithms first perform CCA and continuous wavelet transform of the canonical components to generate a number of features which include component autocorrelation values and wavelet coefficient magnitude values. A subset of the most important features is subsequently selected using RF and labelled observations (supervised case) or synthetic data constructed from the original observations (unsupervised case). The proposed algorithms are evaluated using realistic simulation data as well as 30min epochs of non-invasive EEG recordings obtained from ten patients with epilepsy.

RESULTS

We assessed the performance of the proposed algorithms using classification performance and goodness-of-fit values for noisy and noise-free signal windows. In the simulation study, where the ground truth was known, the proposed algorithms yielded almost perfect performance. In the case of experimental data, where expert marking was performed, the results suggest that both the supervised and unsupervised algorithm versions were able to remove artifacts without affecting noise-free channels considerably, outperforming standard CCA, independent component analysis (ICA) and Lagged Auto-Mutual Information Clustering (LAMIC).

CONCLUSION

The proposed algorithms achieved excellent performance for both simulation and experimental data. Importantly, for the first time to our knowledge, we were able to perform entirely unsupervised artifact removal, i.e. without using already marked noisy data segments, achieving performance that is comparable to the supervised case.

SIGNIFICANCE

Overall, the results suggest that the proposed algorithms yield significant future potential for improving EEG signal quality in research or clinical settings without the need for marking by expert neurophysiologists, EMG signal recording and user visual inspection.

Epileptic seizure onset correlates with long term EEG functional brain network properties

We investigated the correlation of epileptic seizure onset times with long term EEG functional brain network properties. To do so, we constructed binary functional brain networks from long-term, multichannel electroencephalographic data recorded from nine patients with epilepsy. The corresponding network properties were quantified using the average network degree. It was found that the network degree (as well as other network properties such as the network efficiency and clustering coefficient) exhibited large fluctuations over time; however, it also exhibited specific periodic temporal structure over different time scales (1.5hr-24hr periods) that was consistent across subjects. We investigated the correlation of the phases of these network periodicities with the seizure onset by using circular statistics. The results showed that the instantaneous phases of the 3.5hr, 5.5hr, 12hr and 24hr network degree periodic components are not uniformly distributed, suggesting that functional network properties are related to seizure generation and occurrence.

Dynamic cerebral autoregulation in young athletes following concussion

The purpose of this study was to examine cerebral autoregulation (CA) in young athletes experiencing concussion. The subjects were monitored and repeatedly tested 72 hours, 2 weeks and 1 month post-injury. Mean arterial blood pressure (MABP), end-tidal partial pressure of carbon dioxide (PETCO2) and cerebral blood flow velocity (CBFV) in the middle and posterior cerebral arteries were monitored during mental activation paradigms. In order to characterize CA we employed autoregressive models with exogenous inputs (ARX) and impulse response models based on the Laguerre expansion technique (LET). By extracting gain and phase estimates from the obtained models we were able to detect disruptions in CA 72 hours following concussion and a slow recovery within a time period of one month.

Multiple-input nonlinear modelling of cerebral haemodynamics using spontaneous arterial blood pressure, end-tidal CO2 and heart rate measurements

In order to examine the effect of changes in heart rate (HR) upon cerebral perfusion and autoregulation, we include the HR signal recorded from 18 control subjects as a third input in a two-input model of cerebral haemodynamics that has been used previously to quantify the dynamic effects of changes in arterial blood pressure and end-tidal CO2upon cerebral blood flow velocity (CBFV) measured at the middle cerebral arteries via transcranial Doppler ultrasound. It is shown that the inclusion of HR as a third input reduces the output prediction error in a statistically significant manner, which implies that there is a functional connection between HR changes and CBFV. The inclusion of nonlinearities in the model causes further statistically significant reduction of the output prediction error. To achieve this task, we employ the concept of principal dynamic modes (PDMs) that yields dynamic nonlinear models of multi-input systems using relatively short data records. The obtained PDMs suggest model-driven quantitative hypotheses for the role of sympathetic and parasympathetic activity (corresponding to distinct PDMs) in the underlying physiological mechanisms by virtue of their relative contributions to the model output. These relative PDM contributions are subject-specific and, therefore, may be used to assess personalized characteristics for diagnostic purposes.

Spontaneous physiological variability modulates dynamic functional connectivity in resting-state functional magnetic resonance imaging

It is well known that the blood oxygen level-dependent (BOLD) signal measured by functional magnetic resonance imaging (fMRI) is influenced-in addition to neuronal activity-by fluctuations in physiological signals, including arterial CO2, respiration and heart rate/heart rate variability (HR/HRV). Even spontaneous fluctuations of the aforementioned physiological signals have been shown to influence the BOLD fMRI signal in a regionally specific manner. Related to this, estimates of functional connectivity between different brain regions, performed when the subject is at rest, may be confounded by the effects of physiological signal fluctuations. Moreover, resting functional connectivity has been shown to vary with respect to time (dynamic functional connectivity), with the sources of this variation not fully elucidated. In this context, we examine the relation between dynamic functional connectivity patterns and the time-varying properties of simultaneously recorded physiological signals (end-tidal CO2 and HR/HRV) using resting-state fMRI measurements from 12 healthy subjects. The results reveal a modulatory effect of the aforementioned physiological signals on the dynamic resting functional connectivity patterns for a number of resting-state networks (default mode network, somatosensory, visual). By using discrete wavelet decomposition, we also show that these modulation effects are more pronounced in specific frequency bands.

GPU technology as a platform for accelerating physiological systems modeling based on Laguerre-Volterra networks

The use of a GPGPU programming paradigm (running CUDA-enabled algorithms on GPU cards) in biomedical engineering and biology-related applications have shown promising results. GPU acceleration can be used to speedup computation-intensive models, such as the mathematical modeling of biological systems, which often requires the use of nonlinear modeling approaches with a large number of free parameters. In this context, we developed a CUDA-enabled version of a model which implements a nonlinear identification approach that combines basis expansions and polynomial-type networks, termed Laguerre-Volterra networks and can be used in diverse biological applications. The proposed software implementation uses the GPGPU programming paradigm to take advantage of the inherent parallel characteristics of the aforementioned modeling approach to execute the calculations on the GPU card of the host computer system. The initial results of the GPU-based model presented in this work, show performance improvements over the original MATLAB model.

Automatic detection and removal of muscle artifacts from scalp EEG recordings in patients with epilepsy

Automatic detection and removal of muscle artifacts plays an important role in long-term scalp electroencephalography (EEG) monitoring, and muscle artifact detection algorithms have been intensively investigated. This paper proposes an algorithm for automatic muscle artifacts detection and removal using canonical correlation analysis (CCA) and wavelet transform (WT) in epochs from long-term EEG recordings. The proposed method first performs CCA analysis and then conducts wavelet decomposition on the canonical components within a specific frequency range and selects a subset of the wavelet coefficients for subsequent processing. A set of features, including the mean of wavelet coefficients and the canonical component autocorrelation values, are extracted from the above analysis and subsequently used as input in a random forest (RF) classifier. The RF classifier produces a similarity measure between observations and selects a subset of the most important features by comparing the original data with a set of synthetic data that is constructed based on the latter. The RF predictor output is finally used in combination with unsupervised clustering algorithms to discriminate between contaminated and non-contaminated EEG epochs. The proposed method is evaluated in epochs of 30 min from scalp EEG recordings obtained from three patients with epilepsy and yields a sensitivity of 71% and 80%, as well as a specificity of 81% and 85% for k-means and spectral clustering, respectively.

Model-Based Tumor Growth Dynamics and Therapy Response in a Mouse Model of De Novo Carcinogenesis

Tumorigenesis is a complex, multistep process that depends on numerous alterations within the cell and contribution from the surrounding stroma. The ability to model macroscopic tumor evolution with high fidelity may contribute to better predictive tools for designing tumor therapy in the clinic. However, attempts to model tumor growth have mainly been developed and validated using data from xenograft mouse models, which fail to capture important aspects of tumorigenesis including tumor-initiating events and interactions with the immune system. In the present study, we investigate tumor growth and therapy dynamics in a mouse model of de novo carcinogenesis that closely recapitulates tumor initiation, progression and maintenance in vivo. We show that the rate of tumor growth and the effects of therapy are highly variable and mouse specific using a Gompertz model to describe tumor growth and a two-compartment pharmacokinetic/ pharmacodynamic model to describe the effects of therapy in mice treated with 5-FU. We show that inter-mouse growth variability is considerably larger than intra-mouse variability and that there is a correlation between tumor growth and drug kill rates. Our results show that in vivo tumor growth and regression in a double transgenic mouse model are highly variable both within and between subjects and that mathematical models can be used to capture the overall characteristics of this variability. In order for these models to become useful tools in the design of optimal therapy strategies and ultimately in clinical practice, a subject-specific modelling strategy is necessary, rather than approaches that are based on the average behavior of a given subject population which could provide erroneous results.

Periodicity in functional brain networks: application to scalp EEG from epilepsy patients

Seizure detection and prediction studies using scalp- or intracranial-EEG measurements often focus on short-length recordings around the occurrence of the seizure, normally ranging between several seconds and up to a few minutes before and after the event. The underlying assumption in these studies is the presence of a relatively constant EEG activity in the interictal period, that is presumably interrupted by the occurrence of a seizure, at the time the seizure starts or slightly earlier. In this study, we put this assumption under test, by examining long-duration scalp EEG recordings, ranging between 22 and 72 hours, of five patients with epilepsy. For each patient, we construct functional brain networks, by calculating correlations between the scalp electrodes, and examine how these networks vary in time. The results suggest not only that the network varies over time, but it does so in a periodic fashion, with periods ranging between 11 and 25 hours.

Arterial CO2 effects modulate dynamic functional connectivity in resting-state fMRI

It is well known that the blood-oxygen level dependent (BOLD) signal measured by functional magnetic resonance imaging (fMRI) is influenced - in addition to neuronal activity - by fluctuations in physiological signals, including arterial CO2. For instance, even spontaneous CO2 fluctuations have been shown to influence the BOLD fMRI signal regionally. Related to this, studies of functional connectivity between different brain regions, performed when the subject is at rest, may be confounded by the effects of physiological noise. Moreover, resting functional connectivity has been shown to vary with respect to time (dynamic functional connectivity), with the sources of this variation not fully understood at present. In this context, in the present paper we examine the relation between dynamic functional connectivity patterns and the properties of the end-tidal CO2 signal (PETCO2) using resting-state fMRI measurements from 12 healthy subjects. The results demonstrate that there exists a modulatory effect of the correlation strength between PETCO2 and the BOLD signal on dynamic resting functional connectivity. The extent to which this effect was observed was found to be strongly dependent on the data analysis methodology.

Assessment of nonlinear interactions in event-related potentials elicited by stimuli presented at short interstimulus intervals using single-trial data

The recording of brain event-related potentials (ERPs) is a widely used technique to investigate the neural basis of sensory perception and cognitive processing in humans. Due to the low magnitude of ERPs, averaging of several consecutive stimuli is typically employed to enhance the signal to noise ratio (SNR) before subsequent analysis. However, when the temporal interval between two consecutive stimuli is smaller than the latency of the main ERP peaks, i.e., when the stimuli are presented at a fast rate, overlaps between the corresponding ERPs may occur. These overlaps are usually dealt with by assuming that there is a simple additive superposition between the elicited ERPs and consequently performing algebraic waveform subtractions. Here, we test this assumption rigorously by providing a statistical framework that examines the presence of nonlinear additive effects between overlapping ERPs elicited by successive stimuli with short interstimulus intervals (ISIs). The results suggest that there are no nonlinear additive effects due to the time overlap per se but that, for the range of ISIs examined, the second ERP is modulated by the presence of the first stimulus irrespective of whether there is time overlap or not. In other words, two ERPs that overlap in time can still be written as an addition of two ERPs but with the second ERP being different from the first. This difference is also present in the case of nonoverlapping ERPs with short ISIs. The modulation effect elicited on the second ERP by the first stimulus is dependent on the ISI value.

Between-centre variability in transfer function analysis, a widely used method for linear quantification of the dynamic pressure-flow relation: the CARNet study

Transfer function analysis (TFA) is a frequently used method to assess dynamic cerebral autoregulation (CA) using spontaneous oscillations in blood pressure (BP) and cerebral blood flow velocity (CBFV). However, controversies and variations exist in how research groups utilise TFA, causing high variability in interpretation. The objective of this study was to evaluate between-centre variability in TFA outcome metrics. 15 centres analysed the same 70 BP and CBFV datasets from healthy subjects (n=50 rest; n=20 during hypercapnia); 10 additional datasets were computer-generated. Each centre used their in-house TFA methods; however, certain parameters were specified to reduce a priori between-centre variability. Hypercapnia was used to assess discriminatory performance and synthetic data to evaluate effects of parameter settings. Results were analysed using the Mann-Whitney test and logistic regression. A large non-homogeneous variation was found in TFA outcome metrics between the centres. Logistic regression demonstrated that 11 centres were able to distinguish between normal and impaired CA with an AUC>0.85. Further analysis identified TFA settings that are associated with large variation in outcome measures. These results indicate the need for standardisation of TFA settings in order to reduce between-centre variability and to allow accurate comparison between studies. Suggestions on optimal signal processing methods are proposed.

A mixed effects model framework for the assessment of nonlinear interactions in event-related potentials (ERPs) elicited by identical successive stimuli

The recording of brain event-related potentials (ERPs) is a widely used technique to investigate the neural basis of sensory perception and cognitive processing in humans. A commonly used assumption, when dealing with potentially overlapping ERPs elicited by successive stimuli with interstimulus interval (ISI) smaller than the latency of the ERPs, is that their interaction is linear. These overlaps are usually dealt by using averaged waveforms, mostly to enhance the signal-to-noise ratio (SNR) and performing algebraic waveform subtractions. In this paper, we examine the hypothesis of linear interactions by providing a statistical framework that examines the presence of nonlinear additive effects between overlapping ERPs elicited by successive stimuli with short ISIs. The statistical analysis is designed for single trial rather than averaged waveforms. The results suggest that there are no nonlinear additive effects due to the time overlap per se but that, for the range of ISIs examined, the second ERP is modulated by the presence of the first stimulus irrespective of whether there is time overlap or not. In other words, two ERPs that overlap in time can still be written as an addition of two ERPs, with the second ERP being different to the first. The modulation effect on the second ERP by the first stimulus varies for different ISIs.

Multivariate nonstationary modeling of cerebral hemodynamics

We extracted adaptive univariate and multivariate dynamic models of cerebral hemodynamics during resting and hypercapnic conditions using a Recursive Least Squares estimation scheme with multiple adaptive forgetting factors. The time dependent relationship between mean arterial blood pressure (MABP), end-tidal CO2 tension (PETCO2) and middle cerebral artery blood flow velocity (CBFV) was assessed using Laguerre - Volterra models with time varying coefficients. The results suggest that the addition of PETCO2 as a second input yields more accurate and less nonstationary estimates, indicating that unobservable physiological variables are important in the context of nonstationary systems modeling, and particularly for assessing cerebral hemodynamics and autoregulation.

Taking a break from chemotherapy to fight drug-resistance: The cases of cancer and HIV/AIDS

In this work, we present how optimized treatment interruptions during chemotherapy may be used to control drug-resistance, a major challenge for clinicians worldwide. Specifically, we examine resistance in cancer and HIV/AIDS. For each disease, we use mathematical models alongside real data to represent the respective complex biological phenomena and optimal control algorithms to design optimized treatment schedules aiming at controlling disease progression and patient death. In both diseases, it is shown that the key to controlling resistance is the optimal management of the frequency and magnitude of treatment interruptions as a way to facilitate the interplay between the competitive resistant/sensitive strains.

The Volterra-Wiener approach in neuronal modeling

Systems identification is being used increasingly in quantitative neurophysiology, including the auditory, visual and somatosensory systems. In this context, the Volterra-Wiener approach, which is an important branch of nonlinear systems identification, has met with considerable success in neuronal systems modeling, as these systems often exhibit complex nonlinear behavior. The Volterra-Wiener approach provides a comprehensive data-driven framework that does not place any a priori assumptions on the system structure. Therefore, it can approximate highly complex nonlinear mappings provided that experimental protocols are carefully designed in order to meet the requirements of the corresponding estimation procedure. In the present paper, we present a brief overview of Volterra-Wiener models and methodologies for their estimation as they relate to modeling neuronal systems. We also examine a specific example from a mechanoreceptor system.

Recursive least squares estimation of nonlinear multiple-input systems using orthonormal function expansions

We present a computational scheme to obtain adaptive non-linear, multiple-input models of the Volterra-Wiener class, by utilizing function expansions of the Volterra kernels in a recursive least-squares formulation. Function expansions have been proven successful in linear and nonlinear systems identification as they result in a significant reduction of the required free parameters, which is a major limiting factor particularly for nonlinear systems, whereby this number depends exponentially on the nonlinear system order. We illustrate the performance of the proposed scheme by presenting results for a simulated linear two-input system with time-varying characteristics.