ECG Component Classifier Performance

This section shows the results of my proposed ECG component classification method on an arbitrary EEG dataset in automatic mode. For the tests, seven subjects were selected at random from a 61-subject 128-channel closed-eye resting-state EEG dataset.

Table 5-4 lists the performance results obtained on the recordings. For subjects s1 and s2, no sensitivity results could be calculated, since no ECG contamination was detectable in the datasets. Correctly, my method did not find any QRS complexes in the components, and consequently none of the independent components were classified as ECG, meaning, that 𝑇𝑃 = 𝐹𝑃 = 0. For both subjects s1 and s2, no ECG was recognised in the EEG measurements, as the heart effect was so small that it was undetectable.

Dataset Proposed method

QRS detection Sen (%) Classifier Sen (%) Classifier Spe (%)

s1 N/A N/A 100.00

Table 5-4: The ECG artifact detection performance of my proposed method on 128-channel resting-state EEG.

5.4 Summary

In this chapter I proposed a fully automatic ECG artifact removal method working without human assistance or reference ECG channel, which can be used in high-throughout, high-speed EEG analysis, continuous monitoring or clinical diagnostic systems. The acquired EEG signals are subjected to independent component analysis and the resulting independent components are examined for cardiac activity characteristics. The applied adaptive threshold-based QRS detector and subsequent rule-based cardiac cycle classifier identify ECG activity and mark component segments for rejection with high reliability.

In QRS detection, the proposed method achieves sensitivity above 99.3% on the PhysioNet datasets (specificity > 99%), higher than all known automatic methods reported in the literature.

For our high-density resting state EEG data, the QRS detection sensitivity is above 98.1%, however, the sensitivity of the ECG component classifier is 100%. This is due to the fact that the classifier does not need all the component QRS peaks to identify a component segment as ECG.

The significance of my method is that due to its excellent sensitivity and specificity, it can be used reliably for automatic, unsupervised artifact removal, where similar reported methods might incorrectly remove non-artifacts or leave contaminating components in the dataset.

57

My method advances the current practice of ECG artifact removal, and due to its clear advantages, i.e. the fully automatic operation, better sensitivity than previous approaches, and the capability of detecting pathological ECG waveforms, such as frequent ventricular ectopic beats or bundle branch blocks, it will help practitioners in producing more accurate analysis results.

58

6 Functional Connectivity in Ischemic Stroke

As scientific evidence emerged establishing that the operation of the brain is not purely functional (assuming that different parts of the brain are responsible for different well-defined functions) but connectionist (networks of multiple areas execute functions in coordinated cooperation) [139], the focus of research in understanding brain execution mechanisms and brain mapping started to shift from finding individual activated regions to identifying brain networks. When brain imaging and measurement technology advanced the point that it could efficiently support these types of investigations, a new research area, brain connectivity research was born.

As stated in the Introduction, brain connectivity can be divided into three groups as follows:

• structural connectivity: tracks the anatomical fibre pathways between different brain regions [140]. Diffusion Tensor Imaging is used to detect these the interconnecting fibre bundles.

The patterns of these structural connections are relatively persistent at shorter time scales (hours, days), whereas at a longer time scale (months), these patterns may change due to neuroplasticity.

• functional connectivity: is defined as the temporal dependency of the activated neuronal patterns during the flow of information between brain regions. Functional connectivity is based on statistical dependencies between the distant brain regions and can be classified as bivariate or multivariate. The resulting network is represented by undirected graphs.

• effective connectivity: is the measure of causality where one neuronal region has a direct or indirect influence on the activity of another region. This type of connectivity is described by directed graphs.

Connectivity research emerged from MRI technology, first aiming to construct and discover structural pathways (connectome) in the human brain. The emergence of high spatial resolution functional (1-3 mm³) MRI (fMRI) made it possible to conduct functional connectivity studies.

These investigations led to the identification of several fundamental resting-state and task-based brain networks [141–143]. Due to technological limitations, however, fMRI functional connectivity analysis is not suitable for the examination of millisecond range changes typically found, for instance, during cognitive task execution [144]. An alternative to fMRI in connectivity studies is EEG technology that provides superior temporal resolution and measures signals generated directly by neurons as opposed to blood oxygenation changes, such as BOLD fMRI.

EEG functional connectivity can be sensor or source based. If statistical dependence is calculated between the electrode signals, we refer to sensor-level (a.k.a sensor-space) connectivity. If the electrode signals are projected to the cortex by solving the inverse problem that identifies the original sources of bioelectric activities and calculated the association among these cortical regions, we refer to source-level (or source-space) connectivity. Source-level connectivity has the potential to achieve higher spatial resolution (the cortex can be partitioned to thousands of potential source areas) but requires accurate 3D anatomical models and solving the ill-posed inverse problem. For these reasons, sensor-level connectivity would be preferable as an experimental method.

The general process of generating a functional connectivity network from EEG measurements is the following. The cleaned, pre-processed signal is input the first stage of the process that establishes associations between electrodes or cortical sources based on a selected association measure described below. The output of this stage is a square association matrix. Each entry of the matrix represents the strength of the connectivity between two electrodes or sources. This matrix is then used as an adjacency matrix, from which various features can be extracted. To reduce the number of edges in the network graph, normally the association matrix is thresholded

59

and only the top few percent of the edges are kept. The structure of the final connectivity graph can be analyzed by the network features and input to statistical tests.

Functional EEG connectivity has the potential to provide more information than fMRI, due to its higher temporal resolution. Oscillations in the brain regions provide a certain coordination mechanism emerging as synchronized rhythms. These oscillations may transfer information from a local network or region to another region. Examining the flow of information between regions may help to reveal the connectivity relation between the neural assemblies either at rest or during task execution. Connectivity information between the distant brain regions may explain how the neural networks are altered e.g. in stroke or neurodegenerative diseases [145]. It can provide new insights about the large-scale neuronal communication in the brain and may help to understand the origins or track the progress of recovery of stroke or monitor the status of brain diseases such as Alzheimer’s disease [146], and predict outcome of treatment to many other deficits related to the brain. In this chapter, I will focus on how EEG-based functional connectivity can be used to describe brain plasticity in stroke, which is the brain’s natural ability for re-wiring that is essential for successful recovery from stroke.