Fig 1.
Principle of fHR calculation and determination of ST analysis inspired by STAN devices.
Fig 2.
The method works on the principle of adding white noise into the input signal and performing N ensemble trials. The final IMFs are obtained by averaging the results of these trials, the rn(t) is a residual signal.
Fig 3.
Block diagram explaining the ICA-RLS-EEMD method: a) input aECG signals; b) input aECGFIR signals preprocessed by the FIR filter; c) three source components extracted by the ICA method; d) ICA components assigned to the source signals, which were time and amplitude centred and served as inputs to the RLS algorithm; the fECG signal was the output of the ICA-RLS algorithm; e) the first five IMFs that were obtained after the application of the EEMD method; f) a reference scalp electrode recording and the resulting fECG*, which was extracted after applying the ICA-RLS-EEMD method.
Fig 4.
Decomposition of the input aECG signal using the ICA method on three output components (mECG, aECG* and noise): a) inverted polarity of the mECG component; b) mECG polarity correction using the proposed algorithm.
Fig 5.
The RLS algorithm structure with examples of the input and output signals: a) aECG* signal, referred to as primary input or desired signal d(n), b) mECG signal that needed to be adjusted by an adaptive filter, denoted as x(n). Example c) represents an mECGRLS component that has been adjusted by the filter into a shape of the mECG component in the aECG* signal, denoted as y(n). This modified mECGRLS signal was subtracted from the aECG* signal, thus generating the fECG signal, denoted as error signal e(n).
Fig 6.
Averaging of the fQRS complexes and T/QRS calculation of a) recording r01, which achieved high accuracy; b) for recording r04, which achieved poor results.
Fig 7.
Block diagram illustrating the ST segment analysis process.
Table 1.
Statistical evaluation of the fQRS complexes detection obtained by using the ICA-RLS-EEMD method and the FECGDARHA database (the 95% confidence interval is reported in parenthesis).
Table 2.
Statistical evaluation of the fQRS complexes detection obtained by using the ICA-RLS-EEMD method and the Challenge database (the 95% confidence interval is reported in parenthesis).
Table 3.
Summary of records from both databases, which reached threshold values (80% and 95%).
Table 4.
Mean values μ and values of limits of agreement determined by the ICA-RLS-EEMD method for recordings from the FECGDARHA database (the 95% confidence interval is reported in parenthesis).
Table 5.
Mean values μ and values of limits of agreement determined by the ICA-RLS-EEMD method for recordings from the Challenge database (the 95% confidence interval is reported in parenthesis).
Fig 8.
Comparison of reference and estimated values using the ICA-RLS-EEMD method when determining the fHR a) for recording r09 and b) for recording r11 based on the Bland-Altman plots.
Fig 9.
Comparison of reference and estimated values using the ICA-RLS-EEMD method when determining the fHR a) for recording a05 and b) for recording a18 based on the Bland-Altman plots.
Fig 10.
Comparison of fHR traces extracted using the ICA-RLS-EEMD method with annotation a) for all recordings from the FECGDARHA database and b) for 25 recordings from the Challenge database.
Table 6.
Mean values μ and values of limits of agreement determined for ST segment analysis (the 95% confidence interval is reported in parenthesis).
Fig 11.
Comparison of estimated and reference T/QRS ratios of averaged fECG complexes over time.
Table 7.
Comparison of the results with other studies.
Table 8.
Comparison of performance of the methods when extracting the fECG signals from the FECGDARHA database.
Second column provides the number of recordings for which ACC >80% was achieved. Columns three to six show average values calculated for all 12 recordings from the database (the 95% confidence interval is reported in parenthesis).
Table 9.
Comparison of performance of the methods when extracting the fECG signals from the Challenge database.
Second column provides the number of recordings for which ACC >80% was achieved. Columns three to six show average values calculated for 25 recordings from the database (the 95% confidence interval is reported in parenthesis).
Fig 12.
Examples of input aECG signals with different quality and the corresponding outcomes of the ICA-RLS-EEMD method.
Subfigures a) and b) show the examples of the high- (r08) and low-quality (r11) recordings from the FECGDARHA database, for which, respectively, high and low accuracy of the ICA-RLS-EEMD extraction was noticed. Subfigures c) and d) show examples of high- (a15) and low-quality (a18) recordings from the Challenge database, from which fECG was and was not successfully extracted using the ICA-RLS-EEMD, respectively.
Fig 13.
Illustration of the influence of the quality of aECG recordings on the final extraction and the fHR determination.
As an example, the recording r10 from the FECGDARHA database was selected, which was characterized with the overall high-quality extraction. Traces obtained using the ICA-RLS-EEMD method are compared with the annotation. Examples a), b) and c) correspond to sections where high accuracy was achieved in the fHR determination, and examples d), e) and f) correspond to sections where the fHR determination was less accurate.
Fig 14.
Illustration of the influence of the quality of aECG recordings on the final extraction and the fHR determination.
As an example, the recording a16 from the Challenge database was selected, which was characterized by an overall low-quality extraction. Examples a), b) and c) correspond to signal sections where very low accuracy was achieved in the fHR determination, and examples d), e) and f) correspond to sections where the method failed completely when extracting the fHR.
Fig 15.
An example of the effect of the input aECG signal selection on the resulting signal quality: a) optimal selection of input aECG signals (aECG2 and aECG4), the resulting extracted signal is of a high quality; b) inappropriate combination of input signals (aECG2 and aECG3) causing the method to fail to extract fECG signals.
Table 10.
Analysis of the influence of four factors (number of input aECG signals, average ratio between mR and fR oscillations, average value of sSQI and kSQI) on the resulting extraction quality.
The influence of each factor on the accuracy was evaluated using a correlation coefficient (the 95% confidence interval is reported in parenthesis).
Fig 16.
The influence of the mECG and aECG* components determined with the help of the ICA on the resulting quality of the fECG signal extracted using the ICA-RLS method.
The recordings: a) r01 and c) r05 are examples of well extracted mECG and aECG* components. The recording b) r04 and d) r11 represent a less accurate extraction of the fECG. This is a result of the low magnitude of the fetal component and the presence of noise and residues of the fetal component in the mECG signal.
Fig 17.
Influence of the filter order setting on the quality of the extracted fECG signal: a) aECG* and mECG signals used as the inputs to the RLS algorithm; b) low filter order settings, c) optimal filter order settings d) too high filter order settings.
Fig 18.
Illustration of the influence of parameter settings (filter order M and forgetting factor λ) in the RLS method on the resulting quality of the extracted signal a) using the front view 3D graph and b) top-down view.
Fig 19.
The influence of the EEMD parameters (N and Nstd) on the resulting quality of the fECG signal extraction.
Examples a) and b) present the influence of the parameter N, while keeping a constant value of Nstd. Examples c) and d) present the influence of the parameter Nstd, while keeping the N unchanged.
Fig 20.
Illustration of the influence of parameter settings (N and Nstd) for the EEMD method on the resulting quality of the extracted signal a) using the front view of the 3D graph and b) the top-down view.
Fig 21.
Influence of the IMF selection on the final fECG signal quality: a) the optimal selection of IMFs, b) inappropriate selection of IMF2 component representing noise, c) inappropriate selection of both IMFs leading to insufficient suppression of mECG residues, d) inappropriate selection of IMFs, where both signals contain very little information on fECG signal.
Table 11.
Selection of the optimal parameters of the method tested.
Table 12.
Statistical evaluation of the detection of fQRS complexes for the whole recording, while optimizing the method only on the signal section with a length of 15, 30 and 60 seconds (the 95% confidence interval is reported in parenthesis).