Photoplethysmographic determination of the respiratory rate in acutely ill patients: validation of a new algorithm and implementation into a biomedical device

Respiratory rate (RR) is among the first vital signs to change in deteriorating patients, and it has been reported to be the best individual predictor of cardiac arrest in general wards [1]. Current approaches to respiratory rate monitoring may limit their ability to detect changes in patients’ condition [2‐4]. Intermittent RR measurement via visual observation is routine in many patient care settings, and it is usually estimated by visual counts of the chest wall movements for 15 s to 1 min. This approach has several inherent limitations that diminish the clinical utility of these measurements mainly because it is intermittent, and some patients’ “morphology” (e.g. obesity, thoracic deformation) may alter the ability of caregivers to detect chest or abdomen ventilatory-induced movements. Most of all, such non-continuous observation may lead to false appreciation/underestimation of clinical deterioration or clinical worsening.

The photoplethysmographic (PPG) signal is obtained by measuring the intensity of light penetrating through or reflected by the skin, and its widespread application is the routine non-invasive monitoring of arterial oxygen saturation by pulse oximetry, developed in the early 1980s [5, 6]. Recent evidences suggest that RR measurements extracted from PPG may be of clinical interest. The PPG signal includes ventilatory component [7‐12] seen as modulation of the frequency [frequency modulation (FM)] and the amplitude [amplitude modulation (AM)] of the cardiac pulse but also modulation of the intensity of the PPG signal baseline [baseline modulation (BM)] (see “Appendix”). FM is caused by respiratory sinus arrhythmia (RSA), which is a cardiac vagal reflex responsible for a heart rate increase during inspiration and decrease during exhalation [11, 12]. AM is caused by cardiac stroke volume variations resulting from ventilatory influence on venous return, which decreases during exhalation and increases during inspiration in spontaneous breathing (conversely in positive ventilation) [11, 12]. Such ventilatory modulation of the PPG signal is the direct reflection of the respiratory rate [8, 9].

Several studies have attempted to extract RR from PPG recordings [11, 13‐20]. However, the majority of these studies have been carried out in controlled settings using PPG recordings from small groups of healthy subjects, have included PPG waveforms without artefacts (which tends to be frequent in clinical routine), have not depicted results extracted from either a derivation and a prospective validation phase, and were performed at rest and/or in the absence of activity or critical condition such as atrial fibrillation [11, 14‐17, 19, 20]. Most of all, the vast majority of these algorithms have not been implemented in any medical devices for a routine clinical validation, except one [21].

Several other technological approaches for non-invasive RR monitoring techniques have been proposed to address the limitations associated with RR-manual count. These techniques may involve end-tidal CO₂ monitoring, nasal thermistances, respiratory inductance plethysmography via chest bands and ECG-based electrical impedance. However, beside several conflictory results of various RR extractions between these methods, as compared to clinical Ref. [22‐25], the fact that some of these techniques may be considered as intrusive given the dedicated interfaces that are required, the most striking problem for a wide acceptance should be the non-ubiquity of these methods, as compared to PPG that limit their practical application in many clinical settings. While PPG is nowadays ubiquitous in most general wards or step-down facilities, its use to measure RR could facilitate early warning in case of respiratory failure occurrence, thus enabling to monitor concomitantly its two main determinants, i.e. SpO₂ and RR.

The aims of this study were to evaluate the accuracy of respiratory rate measurements using a specifically dedicated reflection-mode PPG signal analysis in a pathological condition and to validate its implementation within medical devices.

Our work presents a feasible option to provide continuous RR monitoring in various types of patients and clinical settings. The observed performance of this technology suggests that it may be a reliable adjunct to pulse oximetry monitoring. To the best of our knowledge, this is the first study to include both a retrospective phase and a prospective validation phase of such a technique, in a wide number of patients and a real-life clinical situation.

Due to the ageing population, respiratory failure is a major public health concern and is burdened with a poor prognosis and a source of prolonged hospitalizations [28]. Early detection and prediction of organ failure (i.e. respiratory failure) could reduce health costs and risks for the patient, offering a reaction early and appropriate medical technology. RR is among the first vital signs to change in deteriorating patients. Adult general ward patients with a RR greater than 24 c/in should be monitored closely and reviewed regularly, even if the other vital signs are normal [29]. A patient with a RR greater than 27 b/min should receive immediate medical review. Intermittent RR measurement via visual observation is routine in many patient care settings; this approach is limited because it is intermittent (could lead to false appreciation of the real patient’s condition by spot check measurements), susceptible to significant human error [30], and requires clinical resources. Transthoracic impedance using ECG electrodes is a popular technique for continuous RR monitoring. This method obtains the respiration signals from the ECG electrodes rather than from the ECG signal. A limitation of thoracic impedance is that a movement artefact introduces errors in RR estimation. It is also sensitive to electrode misplacement and requires constant ECG monitoring which is not ubiquitous in general wards. While used in routine in the operating and recovery room, capnography has also well-documented technical and physiological limitation for spontaneously breathing patients [21], with a high incidence of signal unavailability. However, RR monitoring usually requires biomedical monitoring devices that are not available in clinical routine (ECG continuous monitoring, capnography, acoustimetry, etc.) or human intervention for chronometric measurements. Other continuous monitoring techniques are of limited interest in standard clinical settings either because of the cost and/or availability of the monitoring devices (ex. ECG monitoring or acoustimetry) or because of some technical flaws limiting the availability of the monitoring [31]. The integration of a RR estimation method inside a biomedical technique that is universally used in clinical routine may therefore enable safety improvement. The development of automated RR monitoring modalities may provide more accurate and objective measurements of this vital sign [32]. In such condition, PPG is an interesting approach for ventilation monitoring, as this technique would make simultaneous monitoring of RR and arterial oxygen saturation (pulse oximetry) possible, thus minimizing the number of sensors attached to the patient.

In a study involving 79 patients and volunteers (53 patients and 26 volunteers) aiming to evaluate such a software approach, the authors demonstrated a good correlation between PP RR evaluation and the reference value, with a low bias and limit of agreement [21]. Despite promising results, main limitation of this study was the low number of included patients, the fact that they did not seem to experience acute respiratory failure (mean RR = 16.0 ± 4.6 b/min), and a peak-by-peak comparison, thus limiting discrepancy related to artefacts on continuous measurements. 12.3% of data were excluded from analysis with respect to such interferences, while our results are provided on the entire group of patients and during complete analysis periods, without any data exclusion. The PPG is considered to be very sensitive to any irregularity of the pulse [33]. It means that PPG-RR monitoring may not be available and/or less reliable under certain conditions, such as heart arrhythmia, pacemaker, weak PPG signal and/or excessive movement artefacts.

Considering artefact detection and treatment that were introduced in our algorithm, PPG-RR monitoring remained reliable in 32 patients with atrial fibrillation (20.6 ± 3.7 and 20.5 ± 4.0; R = 0.916) during the retrospective evaluation phase, even if values were not available for some patients. Correlation with reference was high (r = 0.95, p < 0.0001), with a low bias (0.1 b/min) and limit of agreement (96% SD; − 3.3 to + 3.5 b/min), similar to various techniques [21, 31]. However, while remaining within a range of good acceptability (ICC = 0.689) [34], intraclass correlation was lower in case of AF and some patients experienced higher estimation errors. Such results warrant caution before generalization of the method in patients with atrial fibrillation.

The first limit of the study that may be considered is the fact that a contrario to other validation studies for different RR measurement devices and/or algorithms, we chose to compare mean RR values over a clinically relevant period, instead of comparing peak-by-peak values. Such a method was chosen considering that what is really important is the mean RR trend over a significant time period, instead of point-check evaluations; doing this, we also emphasized the potential flaws of our algorithm and increased estimated RR standard deviation, while including periods of time where several artefacts and/or patients’ movements may have increase difficulties to get valid data. We therefore consider that the final device and algorithm were tested in a more clinically relevant situation, instead of doing experimental comparisons solely. Within our algorithm, RR estimation quality was assessed using the standard SpO₂ signal quality index. Further improvement in the estimation accuracy will probably require dedicated artefact detection indexes. A second limit of the proposed method is that, even if quite independent from respiratory drive, PPG-RR monitoring is only reliable when enough evidence on the respiratory component is depicted on the PPG signal during the analysis windows. However, Simoes et al. [35] compared the respiratory rate estimated by manual counting with that obtained using electrical impedance, with chest electrodes. They found that the standard deviation of the difference between respiratory rates estimated using manual counting and that obtained using chest electrodes was 8.6 breaths per minute, with a mean difference of 1.72 breaths per minute. As mentioned earlier, after repeating this calculation for our data, we obtained the same value for standard deviation with a slightly lower mean value. This suggests that the variance of the difference between respiratory rate estimated by manual counting and using electrical impedance measurements using chest electrodes is comparable to the difference between respiratory rates obtained by manual counting and that obtained with our algorithm for processing the PPG waveform. Moreover, even if the limit of agreement of our method was higher in the prospective evaluation phase on 30 patients (96% SD; − 4.5 and + 5.5 c/min), with a low bias (0.5 c/min), it is well below to what is obtained with electrical impedance [31]. Last limitation of the PPG-RR estimation is the fact that it requires a certain level of curves visualization. Exclusion of patients with peripheral vasoconstriction (e.g. hemodynamic instability an unavailability of the PPG curve) may have induced a selection bias and a potential overestimation of the device’s performance.

Our work presents an algorithm for respiratory rate estimation, which uses PPG. PPG is an interesting approach for ventilation and oxygenation monitoring, as this technique would make simultaneous monitoring of respiratory rate and oxygen saturation (pulse oximetry) possible, minimizing the number of sensors attached to the patient in general wards. The algorithm was validated on a great number of patients with various pathologies and a wide range of respiratory rates. It can be used as a reliable adjunct to pulse oximetry monitoring, in order to assess patients’ clinical status.