Automated Detection and Removal of Cardiac and Pulse Interferences from Neonatal EEG Signals

Figures

• 

  Figure 1. Flowchart of the complete data processing for the automated classification of cardiac-related interference ICs. The complete data-processing pipeline includes the sequential stages of data-processing and their respective modular processing steps, input and output.

• 

  Figure 2. Exemplary waveforms of an artificial pulse signal (a) and an ECG signal (b) in a time interval corresponding to one heartbeat period.

• 

  Figure 3. Examples of a real ECG signal synchronously acquired with the EEG signals and the corresponding APS. (a) Time course of a real ECG signal (3 s). (b) Time course of the APS corresponding to ECG shown in (a). (c) Time course of the superimposed real ECG (red line) and APS (black line).

• 

  Figure 4. Flux diagram of the automated process for the classification of the artefactual component as PCC (a) and ECC (b) through a cascade of three criteria. Acronyms: PSDcorrP: correlation value between the component PSD and the APS PSD; TCcorrP: correlation value between the time course of the component and the time course of APS; PSP: Phase Shift value between the component and APS; PSDcorrE: correlation value between the component PSD and the ECG PSD; TCcorrE: correlation value between the time course of the component and the time course of real ECG; PSE: Phase Shift value between the component and the ECG.

• 

  Figure 5. Exemplary components classified as artefactual in the three signal groups (i.e., decomposition to 21, 20 and 19 ICs for G1, G2 and G3, respectively). Examples refer to the same EEG segment. Scalp topography, 3 s of time course and PSD of one IC classified ECC (a) and of two ICs classified as PCC (b,c) are shown for the decomposition to 21 ICs. The weights of the mixing matrix corresponding to the nECG and nAPS signals are represented by circles (pulse activity on the left, heart activity on the right). Scalp topography, 3 s of time course and PSD of one IC classified as ECC (d) and one IC classified as PCC (e) are shown for the decomposition to 20 ICs. The weights of the mixing matrix corresponding to the nECG is represented by a circle. In (f), the scalp topography, 3 s of the time course and PSD of one IC classified as PCC is shown for the decomposition to 19 ICs.

• 

  Figure 6. Examples of cardiac artefact removal in three EEG channels (frontal: F8, central: C4 and temporal: T8). Panels (a,c,e) show 2 s of the time course of three exemplary EEG signals from one neonatal EEG segment. Panels (b,d,f) show the PSD around the specific heart rate frequency (i.e., 2.23 Hz). The blue lines indicate the PSD and time course before the removal of cardiac-related interference. The red, black and green lines indicate the PSD and time course after removal of the cardiac-related interference using the three signal groups, i.e., decomposing to 21 ICs, 20 ICs and 19 ICs.

• 

  Figure 7. Distribution of CC and API for the different decompositions. Box plots of the percentage of contaminated ICs (CC, (a)) and of the Artefactual Percentage Index (API, (b)) for the three signal groups (G1, 21 ICs: blue, G2, 20 ICs: orange; G3, 19 ICs: grey). The horizontal black lines, the boxes, the vertical bars, and the dots indicate the median, the 25–75 percentile, the 5–95 percentile and the outliers, respectively. Single and double asterisks indicate significant differences between signal groups with p < 0.05 and p < 0.001, respectively, as assessed by the Wilcoxon test.

• 

  Figure 8. Distribution of SPV, varRPP and varPSDhr for the different decompositions. Box plots of percentage of SPV (a), varRPP (b) and varPSDhr (c) for the three signal groups (21 ICs: blue, 20 ICs: orange; 19 ICs: grey). The horizontal black lines, the boxes, the vertical bars, and the dots indicate the median, the 25–75 percentile, the 5–95 percentile and the outliers, respectively. Single and double asterisks indicate significant differences between the different signal groups with p < 0.05 and p < 0.001, respectively, as assessed by the Wilcoxon test.

• 

  Figure 9. Examples of cardiac artefact removal in two EEG segments from one channel (T7) of an EEG recording containing seizures (G1s). Panel (a) shows 2 s of the time course of the EEG signal where no seizures were annotated, whereas panel (c) shows 2 s of the time course of the EEG signal during seizure. Panels (b,d) show the PSD around the specific heart rate frequency (i.e., 1.99 Hz (b) and 1.95 Hz (d)). The blue lines indicate the time course and the PSD before the removal of the cardiac-related interference. The red lines indicate the time course and the PSD after removal of the cardiac-related interference (decomposition into 21 ICs).

• 

  Figure 10. Distribution of CC and API for the group G1s in comparison to group G1 (both groups separated into 21 ICs). Box plots of the percentage of contaminated ICs (CC_S, in light blue, refers to G1s; CC, in light yellow, refers to G1) and of the Artefactual Percentage Index (API_S, in blue, refers to G1s and API, in yellow, refers to G1). The horizontal black lines, the boxes, the vertical bars, and the dots indicate the median, the 25–75 percentile, the 5–95 percentile, and the outliers, respectively.

• 

  Figure 11. Distribution of SPV, varRPP and varPSDhr for the group G1s in comparison to group G1 (both groups separated into 21 ICs). Box plots of percentage of SPV (blue for G1s and light blue for G1), varRPP (dark grey for G1s and light grey for G1) and varPSDhr (dark green for G1s and light green for G1). The horizontal black lines, the boxes, the vertical bars, and the dots indicate the median, the 25–75 percentile, the 5–95 percentile and the outliers, respectively. Single and double asterisks indicate significant differences between the G1s and G1 with p < 0.05 and p < 0.001, respectively, as assessed by the Wilcoxon test.