Information Retrieval from Photoplethysmographic Sensors: A Comprehensive Comparison of Practical Interpolation and Breath-Extraction Techniques at Different Sampling Rates

Figures

• 

  Figure 1. Application of the ENVELOPE (a) and SLOPE (b) beat detection methods to the respective signals of interest. (a) On the left, derivative of the PPG signal (grey) and relative superior and inferior envelopes (black); on the right, PPG derivative after subtraction of the inferior and superior envelopes (grey), threshold for peak detection (black line), and detected peaks (red circles). (b) On the left, the original PPG signal (light blue) and the computed adaptive threshold (black); on the right, detected peaks (red circles).

• 

  Figure 2. Application of the described interpolation strategies (a: SPLINE; b: PARABOLA; c: REG) to the peaks identified through the ENVELOPE method. The first derivative of a PPG signal sampled at 32 Hz is shown (blue dots). The following is in red: spline-interpolated first derivative (a), parabola approximation (b), regression curves estimated on the ascending and descending segments (c). The plots show the peaks detected on the subsampled signal (O) and those identified through interpolation (X).

• 

  Figure 3. Application of the FILT (a), ENVL (b), and INTR (c) methods to the respective signals of interest. On the first column, input signals and main features are: (a) PPG signal (light blue); (b) derivative of the PPG signal (grey) and relative superior envelope (black); (c) PPG signal (light blue) and points of maximum slope (red). Second column: breath signals obtained with each method.

• 

  Figure 4. Comparison of the error variance measured on the IBIs with the theoretical one as a function of the sampling frequency. Without PPG interpolation (black points), the measured error variance resembles the expected power law. With interpolation (triangles, circles, and crosses), the error variance measured with the different methods almost overlaps from 64 to 32 Hz; at 16 Hz, an increase in error variance is observed with the REG method, whereas PARABOLA and SPLINE values remain overlapped.

• 

  Figure 5. Boxplots of the beat-to-beat AEs computed with the original IBIs and with those obtained through SPLINE and PARABOLA interpolations. Significant differences (Bonferroni-corrected comparisons, p < 0.05) are highlighted between AEs within each sampling frequency and protocol condition. a: AEs related to SPLINE (S) or PARABOLA (P) are significantly different from the original; b: AEs related to SPLINE (S) are significantly different from PARABOLA.

• 

  Figure 6. MAE (bars), upper, and lower quartiles (error bars) of the considered time- and frequency-domain features. R, S, and P indicate frequencies at which significant differences (Bonferroni-corrected, p < 0.05) were found between the original IBIs and REG (R), SPLINE (S), or PARABOLA (P) interpolations. All the reported differences show effect size (Cohen’s r) > 0.5. *: Significant difference revealed after excluding REG.

• 

  Figure 7. (a): average PSDs of the original PPGs (sampling frequency: 256 Hz, bandwidth: 0–64 Hz) collected from all the participants; (b): cumulative contribution of the different frequency components to total power (y-axis starts at 85% to highlight differences among phases). Columns represent different protocol phases. Each participant’s PSD was computed through Welch’s periodogram after standardizing the PPG signal with z-score to reduce inter-subject variability.

• 

  Figure 8. Example of PSD and magnitude-squared coherence computed for a single subject with the examined breathing signal estimation methods from a 64 Hz PPG signal. The shaded area in the coherence charts indicates the relevant frequency band selected for the participant.