Continuous Patient-Independent Estimation of Respiratory Rate and Blood Pressure Using Robust Spectro-Temporal Features Derived From Photoplethysmogram Only

Figures

• 

  Fig. 1.. PPG signal with arbitrary units (a.u.) showing the respiratory Induced Variations (RIVs), where RIIV is the variation in beat peaks (Envelope). RIFV is the changes in the inter-beat interval. RIAV is the change in the beat strength.

• 

  Fig. 2.. Ideal PPG Morphology exhibiting morphological features for Blood Pressure Estimation.

• 

  Fig. 3.. Output of IMS Algorithm showing peaks and onsets of valid beats. The remaining beats are artifactual having clipped or saturated pulses.

• 

  Fig. 4.. Proposed approach. A: Baseline shifts of raw PPG are removed, and artifacts are identified. B: Physio, spectral and statistical features are extracted from cleaned PPG. C: Extracted features are checked for their relative significance in determining RR and BP. D: Finally, RR is estimated using breath counting, and BP (Systolic and Diastolic) is measured using a DNN regression model.

• 

  Fig. 5.. Three PPG signals with different morphologies from the MIMIC Database with (a) ideal morphology with visible critical points such as inflection point, dicrotic notch, and systolic and diastolic peaks, (b) and (c) nonideal shapes of PPG morphology due to missing or altered peaks/points such as inflection points, dicrotic notch, and diastolic peak.

• 

  Fig. 6.. 32 sec PPG window displaying Physiological Feature Extraction for RR estimation: RIIV, RIFV, and RIAV are formed using identified peaks and onsets, and then re-sampled to 4 Hz. RIIV is the most accurate representation of respiration activity.

• 

  Fig. 7.. Morphological features extracted from PPG morphology for BP estimation.

• 

  Fig. 8.. PPG Spectrum showing 5 frequency bands from 0 to 10 Hz for Spectral Feature Extraction.

• 

  Fig. 9.. Examples of two 32 sec extracted modulation window (left) and modulation power spectrum (right) where (a) shows that the respiration frequency is strongly present in the modulation, and (b) shows that the respiration frequency is weakly present in the modulation. The dotted blue margins indicate a plausible respiration range of 0.1 to 1 Hz (6 to 60 bpm) while red ones show 3 power frequency components middling the peak power value.

• 

  Fig. 10.. Examples 32 sec window of extracted modulation (top) and templates (bottom) where (a) represents a good quality waveform having a strong correlation with the template, and (b) represents bad quality modulation having less \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${MQI}_{XC-Tmp}$\end{document}.

• 

  Fig. 11.. Effect of changing Thresholds on (a) Performance measured through mean absolute error (bpm) and (b) Data Retention measured through % of windows discarded. As the threshold value is increased (strict quality thresholds), Error decreases at the expense of discarding many data windows. The red circle indicates the median and blue line boundaries denote the Inter-Quartile Range (25th to 75th percentiles).

• 

  Fig. 12.. Neural Network Architecture for BP estimation.