Calibration for massive physiological signal collection in hospital -- Sawtooth artifact in beat-to-beat pulse transit time measured from patient monitor data

Objective: Calibration is one of the most important initial steps in any signal acquisition and experiment. It is however less discussed when massively collecting physiological signals in clinical setting. Here we test an off-the-shelf integrated Photoplethysmography (PPG) and electrocardiogram (ECG) monitoring device for its ability to yield a stable Pulse transit time (PTT) signal. Method: This is a retrospective clinical study using two databases: one containing 35 subjects who underwent laparoscopic cholecystectomy, another containing 22 subjects who underwent spontaneous breathing test in the intensive care unit. All data sets include recordings of PPG and ECG using a commonly deployed patient monitor. We calculated the PTT signal offline. Result: We identify a novel constant oscillatory pattern in the PTT signal and identify this pattern as the sawtooth artifact. We propose an approach based on the de-shape method to visualize, quantify and validate this sawtooth artifact. Conclusion: The PTT and ECG signals not designed for the PTT evaluation may contain unwanted artifacts. The PTT signal should be calibrated before analysis to avoid erroneous interpretation of its physiological meaning.


Introduction
Calibration is one of the most important initial steps in any signal acquisition and experiment ----the data collection equipment, or the quality of the data, needs to be calibrated before a meaningful data analysis can take place. By calibration, we mean the validity of the signal resource, and checking if the signal is correctly recorded for the specific purpose. In our high--volume data fever era, we rely heavier than ever before on "big data" analysis [Raghupathi2014], and it is popular now to collect as much data as possible in the hospital research setting, as for example from a patient monitor and Holter ECG. While there have been a lot of discussion about the artifact issues in patient monitoring data [Takla2006,Nizami2013,Hravnak2016,Chen2016], how the data is calibrated, or if the data can be applied for a specific purpose, is often not discussed and, rather, implicitly assumed. Particularly, when multiple time series recorded from the marketed patient monitor are analyzed in the framework of sensor fusion [Gravina2017], it is often implicitly assumed that on the machine level the relationship between channels, like synchronization, is not a problem. More generally, it is not well confirmed which online available datasets are suitable for which purposes, since the original recording machine might not be designed for the intended purpose. Note that this calibration issue is one source of artifacts in patient monitor data, but to the best of our knowledge, it has not been discussed in the society. In this paper, we provide an evidence showing the potential problem when analyzing databases massively collected from the hospital environment without properly taking this calibration step into account. We demonstrate that an artifact in the pulse transit time (PTT) signal [Mukkamala2015] referred to as sawtooth artifact can occur because the marketed patient monitor was not designed and calibrated for this specific purpose in the first place. In particular, the calculation of PTT requires a precise time measure of both R peak of ECG and the pulse wave arrival of PPG. Unfortunately, the PTT derived from the recorded photoplethysmography (PPG) and electrocardiogram (ECG) signals from a marketed patient monitor is not precise enough in some devices. Therefore, the subsequent calculated PTT contains a regularly wandering pattern of pulse wave arrival time, producing the appearance of sawtooth.
We observed this sawtooth artifact in a clinical data set exemplified in Figure 1 showing a recording during the entire laparoscopic cholecystectomy (LC) procedure.
Although both PTT calculated from arterial blood pressure and PTT calculated from PPG demonstrate a change that is reciprocal to the blood pressure on the large scale, PTT calculated from PPG contains the sawtooth oscillation with almost constant periods. To evaluate systematically if this sawtooth oscillation is physiological or artifact, and if it is consistent across different subjects, we carried out the present study. Below, we focus on PTT calculated from PPG.

Materials
The data set used in the present manuscript comes from two prospective observational studies.
The first study has been approved by the local institutional ethics review boards (Shin Kong Wu Ho--Su Memorial Hospital, Taipei, Taiwan; IRB No.: 20160706R). Written informed consent was obtained from each patient. From Dec. 2016 to Oct. 2017, we enrolled 33 patients, ASA I to III, scheduled for LC. Inclusion criteria were patients with acute cholecystitis, chronic cholecystitis or gall stone eligible to undergo LC surgery. All surgeries were carried out by one surgeon to ensure the consistency of surgical procedures across all cases. Exclusion criteria were major cardiac problems, uncontrolled hypertension, arrhythmia shown in pre--operative ECG, known neurological disease, history of drug abuse and anticipated difficult airways.
The recording lasted on average 26.9 minutes with standard deviation of 5.25 minutes and captured the surgical steps from laparoscopic ports establishment to major part of the gallbladder removal. The physiological data including PPG and ECG were simultaneously recorded from two Philips IntelliVue MP60 Patient Monitors and one Philips IntelliVue MX800 Patient Monitor. Three machines installed in three different operating rooms were used throughout the study. These patient monitors were serviced by an engineer of the manufacturer on quarterly basis. The data were collected via data dumping system provided by the third--party software, ixTrend Express ver. 2.1 (ixellence GmbH, Wildau, Germany. https://www.ixellence.com/index.php/en/products/ixtrend). The sampling rates of ECG (lead II in EASI mode) and PPG channels were 500 Hz and 125 Hz, respectively. We refer to this first database as SKWHSMH database.
The second study has been approved by the local institutional ethics review boards (Chang Gang Memorial Hospital, Taipei, Taiwan; IRB No.:104--6531B). Written informed consent was obtained from each patient. From Nov. 2016 to Feb. 2017, we enrolled 22 patients in the medical intensive unit who were ready for weaning. Inclusion criteria were patients that have been intubated for longer than 24 hours, and ready for weaning. Specifically, the patient showed clear improvement of the condition which led to mechanical ventilation; acute pulmonary or neuromuscular disease or increased intracranial pressure signs were not present; consciousness and semi--recumbency were required; PaO2 ≥ 60mmHg and FiO2 ≤ 40% with PEEP ≤ 8cm H2O, or PaO2/FiO2 >150 mmHg; PaCO2<50mmHg or increasing <10% for patients with chronic CO2 retention; Heart rate <140 bpm and the systolic blood pressure of 90--160mmHg; no vasopressive or inotropic drugs administered for more than 8 hours; no intravenous sedation within the previous 24 hours; ability for the patient to cough while being suctioned; afebrile with the body temperature less than and equal to 38• C; negative cuff leakage test >110ml or >12%. Exclusion criteria were the presence of a tracheostomy, or the patient having been on home ventilation prior to ICU admission, or the patient's or family's decision not to re--intubate or withdrawal from the care anticipated, or planned surgery requiring sedation within the next 48 hours.
All recordings lasted 5 minutes during the spontaneous breathing test. The physiological data including PPG and ECG were simultaneously recorded from several Philips IntelliVue MP60 Patient from different beds. These patient monitors were also serviced by an engineer of the manufacturer on quarterly basis. The data were collected via data dumping system provided by the third--party software, MediCollector ver. 1.0.46 (MediCollector, USA. https://www.medicollector.com). The sampling rates of ECG (lead II in EASI mode) and PPG channels were 500 Hz and 125 Hz, respectively. We refer to this second database as CGMH database.
For the reproducibility purpose, the ECG signal, the PPG signal, and the derived PTT signal of the two databases are made publicly available in doi:10.7910/DVN/OJBZ67.

Methods
We resampled the PPG to 500 Hz by using linear interpolation method for PTT calculation. To avoid any possible issue caused by the data dumping system, the quality of the dumped signal was visually compared with the signal displayed on the patient monitor. The entire period of the recorded signal was analyzed as described below.
We followed the method of PTT calculation previously reported [Gesche2012]. The time point of R--peak was determined from the lead II ECG signal. The time point of pulse wave arrival was determined by the maxima of the first derivation during the ascent of the waveform; that is, the location of the fastest ascending PPG waveform. The time interval between each pair of R--peak time and subsequent pulse wave arrival time was calculated and resampled at 4 Hz by the cubic spline interpolation. We refer to the PTT calculated from ECG and PPG as PTT PPG , and that calculated from ECG and invasive arterial blood pressure waveform as PTT ABP . To further quantify how the PTT oscillates, particularly locally, and evaluate if the oscillation is physiological, we applied the recently developed de--shape short--time Fourier transform algorithm (dsSTFT) [Lin2018]. Compared with the other time-frequency (TF) representations, dsSTFT provides a nonlinear--type TF representation that shows only the fundamental instantaneous frequency of the oscillatory signal [Lin2018]. We chose it since the oscillatory pattern in PTT is non--sinusoidal, and most other TF representations will be complicated by the inevitable multiples. For the reproducibility purpose, the dsSTFT code can be downloaded from https://hautiengwu.wordpress.com/code/.

Results
For the SKWHSMH database, out of 33 cases, in 11, 19, 2, and 1 subjects, we took a 2000s, 1500s, 1200s and 500s--period of data, respectively, after the stabilization of PPG, ECG and ABP signals, as determined by visual inspection of the waveforms. The unified sizes simplified the observation of the sawtooth artifact. Except one subject, the selected periods covered the majority of the surgical period in each case. The PTT signals of all subjects and their associated power spectra are shown in Figure  2. It is clear that in all cases, there is a clear sawtooth oscillation in the PTT signal. In the associated power spectra, dominant peaks around 0.01 Hz, 0.012 Hz and 0.1 Hz are observed. Among 33 cases, there are 10 cases with the dominant frequency at 0.01 Hz, 15 cases with the dominant frequency at 0.012 Hz, and 8 cases with the dominant frequency slightly below 0.1 Hz. The TF representation of the PTT signal by dsSTFT is shown in Figure 3. In each plot, the x--axis denotes the time in seconds, the y--axis is the frequency in Hertz (Hz), and the intensity of the image means the strength of the oscillation inside the PTT at each time and frequency. From the TF representation, there is a dominant line at 0.01 Hz from the beginning to the end in the demonstrated subject. This indicates that at each moment, the PTT signal shows a regular oscillation at 100 seconds period. Coming back to the time series, we see that the artifact is not only oscillating at 0.01 Hz at each moment, but also with the sawtooth pattern. This artifact is the same as those shown in Figure 1. This local 0.01 Hz oscillation shows up in 10 subjects and the frequency is fixed and persists throughout the LC procedure and regardless of the surgical and anesthetic manipulations. For the other subjects, a local 0.1 Hz oscillation was observed (not shown). To illustrate this fact for all 35 subjects, the mean TF representations of all 35 subjects are shown in the bottom subplot of Figure 2. Although the whole signal was analyzed, for the visualization purpose, only the first 1,600 seconds are shown here, since 1,600 seconds is the shortest recorded signal across all recordings. It is clear that the only dominant curve left after taking average is again the 0.01 Hz curve. Note that although all subjects received LC, the timestamps of different intrasurgical interventions varied. Even under this heterogeneous situation, the 0.01 Hz oscillation persisted. This indicates that this sawtooth oscillation is common across all subjects, which makes it unlikely to be physiological. For the CGMH database, we made similar observations. The PTT signals of all subjects and their associated power spectra are shown in Figure 4. There is a sawtooth oscillation in the PTT signal in all cases. In the associated power spectra, there are 9 cases with the dominant frequency at 0.01 Hz (100 second period), 4 cases with the dominant frequency at 0.0067 Hz (150 seconds period), 5 cases with the dominant frequency 0.0133 Hz (75 seconds period), 2 cases with the dominant frequency at 0.1Hz (10 second period), and 2 cases with other dominant frequencies.
The TF representation of the PTT signal by dsSTFT is shown in Figure 5. Again, we can see a dominant 0.01 Hz in the averaged TF representation. Note that compared with those in Figure 3, the plot is more blurred. It is because the PTT signal in the CGMH database is much shorter. In a 300 seconds period, the artifact of 100 seconds period only appears three times, which degrades the TF representation quality.

Conclusion
The main finding of our study is the possible erroneous information from massively collected data from the hospital when the machine is not designed for the intended purpose. Specifically, we provide an evidence from the PTT signal extracted from a clinically widely used monitoring machine that, albeit enabling PTT analysis, was not designed for this purpose. An example of such information is the sawtooth artifact presenting as a non--physiological oscillation in the PTT signal as we have shown in the present investigation. While we could not systematically examine all off--the--shelf monitoring devices on the market, we suspect that a similar issue might exist in other machines, other combinations of different channels or may present in other formats. To the best of our knowledge, this is the first observation of such measurement artifact. PPG is ubiquitous in medicine. This optics--based noninvasive signal acquisition technique provides a continuous and convenient display of arterial pulse in finger or in earlobe, and the reading of oxygenation by pulse oximetry. The display of peripheral pulse also allows the assessment of heart and respiratory rates [Shelley2007, Tang2017]. When combined with other physiological signals, PPG provides an even wider spectrum of applications in healthcare. For example, PPG amplitude combined with pulse rate helps the assessment of surgical stress during anesthesia [Huiku2007]. The addition of ECG helps the adjustment of a cardiac pacemaker, identification and classification of cardiac arrhythmia [Tang2017]. PTT is an important application of the combination of ECG and PPG, which can be used as a surrogate of pulse wave velocity to indirectly measure the blood pressure [Gesche2012, Kim2013]. Since the standard patient monitoring instruments that are commonly used in clinical anesthesia and critical care medicine are equipped with ECG and PPG, it is intuitive to measure the PTT by calculating the data exported from the monitor to identify novel predictive features. The observed sawtooth artifact, if not being noticed beforehand, might be over--interpreted with a misleading conclusion. What is the most likely underlying cause for the observed artifact? According to a private communication with a Philips engineer, the patient monitor is not designed for the PTT analysis. However, since the design details of the patient monitor and third--party software algorithms are not accessible to us, we do not have a concrete answer of how it happens. However, this finding is connected to a larger issue that is intimately related to data analysis in general: the calibration of the data acquisition procedure. To efficiently use the massively collected physiological signals from generic medical equipment for the non--intended use, we need to confirm if the physiological signal is suitable for the purpose. To the best of our knowledge, while it is generally considered to collect as much data as possible in the hospital, this calibration issue is often not discussed and, rather, implicitly assumed. The demonstrated sawtooth artifact might mislead study conclusion, or even the following steps. Note that while the sawtooth artifact in the demonstrated databases is relatively easily identified, it is conceivable to encounter other, more difficult--to-track artifacts in other setups. We mention that if needed, we can apply signal analysis tools to remove the sawtooth artifact, and hence maximize the utilization of available physiological signals [Lu2018]; for example, the manifold learning tool proposed in [Su2017]. See Figure 6 for an example of the artifact removal based on the manifold learning tool. However, while this post--processing could help us maximize the utilization of currently available data, its reliability needs to be further evaluated before it can be applied. Since it does not solve the original artifact problem, we do not extensively discuss it in this paper.
In conclusion, we demonstrate a negative finding about PTT coming out of a patient monitor that is not designed for the purpose of PTT analysis, even if these channels have been integrated into a single machine. Researchers should pay attention when they analyze signals from machine that is note designed for the intended purpose. If the machine is not designed for a specific purpose, a calibration is needed to confirm the quality of the signal for the intended purpose. Without a suitable calibration, the collected signal might contain unexpected, and the data analysis itself might not be meaningful, or even harmful if misleading conclusions are to be drawn.     a.u. Figure 5. The time--frequency (TF) representation of the PTT (blue curve) determined by the de--shape short--time Fourier transform of subjects from the CGMH database. Left figure shows the PTT from a subject and its TF representation. Right figure shows the average TF representation over 22 subjects. It is clear that there is a dominant line at 0.01Hz. This indicates a regular oscillation of 0.01Hz in most PTT signals. Figure 6. An illustration of removing the sawtooth artifact by the manifold learning tool. The black tracking is the original PTT signal, while the blue one is the corrected PTT signal.