Artifacts classification and apnea events detection in neck photoplethysmography signals

Figures

• 

  Fig. 1. Example of neck and finger PPG pulse waveforms

• 

  Fig. 2. Classification pipeline. The predicted output classes for artifacts classification are: Artifacts/clean PPG; for apnea classification: Apnea/Normal PPG. This process was repeated 30 times with different randomization in the data partition stage, for each window length and threshold of corruption (%) combination

• 

  Fig. 3. Data partitions for artifacts classification and apnea detection models. (a) Leave-30%-of-Subjects-Out approach for artifacts classification. (b) Leave-1/3-of-Recording-Out per subject for apnea detection

• 

  Fig. 4. Average classification results for the proposed artifacts and apnea classification algorithms, over the 30 randomization experiments. Bar graphs show the average performance metrics across the different windows and corruption thresholds (%) tested. The error bars represent the extent of the standard deviation above and below the mean. Different thresholds of corruption (Thd = 20%, 30%, 40%, 50%) are specified as separate coloured bars for each window length (W = 4,5,6,7,8,10s). The statistical results of the multiple pairwise comparisons testing for the window effect are displayed with a horizontal line and a black asterisk symbol indicating the alpha significance level: * p < 0.05, ** p < 0.01, *** p < 0.001. The red asterisks on top of some window groups indicate that all the multiple comparisons were statistically significant for that specific window

• 

  Fig. 5. Classification decision results of the best models for one head movement artifact and an apnea event. The variation of some of the features used are displayed in the lower panels: Peak Height Difference, the maximum of the envelope and the Spectral Entropy (< 1.5Hz). True artifacts and apnea windows are labelled in red and blue respectively

• 

  Fig. 6. Means plots of the classification performance metrics across different thresholds of corruption (%). The means, with the corresponding 95% confidence intervals, are represented in red. Standard deviations above and below the mean are shown in blue. The statistical results of the multiple pairwise comparisons testing for the threshold effect are displayed with a horizontal line and a black asterisk for different alpha significance levels: * p < 0.05, ** p < 0.01 *** p < 0.001. The red asterisks on top of some threshold groups indicate that all the multiple comparisons were statistically significant for that specific window

• 

  Fig. 7. Features selection ranked by frequency of occurrence over the 30 randomization experiments for the best artifacts and apnea classification models. F1 = Amplitude, F2 = Width, F3 = PeakHeightDiff, F4 = PeakDistance, F5 = TroughDiff, F6 = RiseTime, F7 = Skewness, F8 = Kurtosis, F9 = ChangeAmplitude, F10 = ChangeWidth, F11 = ChangePeakHeightDiff, F12 = ChangePeakDistance, F13 = ChangeTroughDiff, F14 = ChangeRiseTime, F15 = ChangeSkewness, F16 = ChangeKurtosis, F17 = StdAmplitude, F18 = StdWidth, F19 = StdPeakHeightDiff, F20 = StdPeakDistance, F21 = StdTroughDiff, F22 = StdRiseTime, F23 = StdSkewness, F24 = StdKurtosis, F25 = ZeroCrossingRate (ZCR), F26 = CorrelogramPeak1, F27 = CorrelogramPeak2, F28 = CorrelogramLag1, F29 = CorrelogramLag2, F30 = SpectralEntropy (0–1.5Hz), F31 = SpectralEntropy (1–4Hz), F32 = SpectralKurtosis (0–1.5Hz), F33 = SpectralKurtosis (1–4Hz), F34 = RelativePower (0–0.8Hz), F35 = RelativePower (0.8–1.3Hz), F36 = RelativePower (1.3–1.8Hz), F37 = AvgPower (0–0.8Hz), F38 = AvgPower (0.8–1.3Hz), F39 = AvgPower (1.3–1.8Hz), F40 = AvgPower (2.2–2.8Hz), F41 = AvgPower (3.2–3.8Hz), F42 = EnvelopeStd, F43 = EnvelopeMax, F44 = EnvelopeMin, F45 = EnvelopeRange, F46 =…