2.1 Study design and population
For this prospective multicentric study, we obtained informed consent from 121 patients aged 18 years or more and scheduled for an elective surgery requiring general anesthesia and invasive BP monitoring at CHUV (University Hospital of Lausanne, Switzerland) and HUG (University Hospital of Geneva, Switzerland) between April 2019 and November 2019. Exclusion criteria were patient refusal, age below eighteen, inability to give informed consent, ASA risk > 3, dysrhythmia (bigeminy, trigeminy, isolated VPB, atrial fibrillation), contraindication to the placement of an arterial catheter (Raynaud’s disease, Burger arthritis, negative Allen test, major hyperlipidemia), or known contact dermatitis to nickel/chromium [
26].
Patients were prepared for anaesthesia according to the existing safety and standard procedures of the Department of Anesthesiology of CHUV Lausanne and HUG Geneva, tailored individually to the patient, depending on his concomitant disease, treatments, and procedures. A dedicated catheter (BD Arterial Cannula 20G/1.1 mm × 45 mm, Becton Dickinson Infusion Therapy Syst. Inc., UT, USA) was inserted predominantly into the right or left radial artery under local anesthesia, allowing beat-to-beat continuous blood pressure monitoring. Patients were monitored with Philips® IntelliVue MP50 monitor conjointly with M3001A module for IAP measurements (Philips, Amsterdam, the Netherlands). General anesthesia was induced with an infusion of propofol (2–3 mg/kg) while intravenous analgesia was provided by boluses of fentanyl (1–2 μg/kg) or continuous remifentanil infusion (0.1–0.5 μg/kg/min) depending on the surgical intervention. Rocuronium (0.6 mg/kg) was administered before intubation. The management of the anesthesia was left at the discretion of the anesthesiologist in charge and maintenance provided with propofol (6–12 mg/kg/h). Boluses of ephedrine (5–10 μg) or phenylephrine (50–100 μg) depending on the clinical context were used in case of necessity for hemodynamic support.
The study was approved by the local ethics committee (CER-VD no 2018-01656) and registered under number NCT03875248 (Arm 1) at
www.clinicaltrials.gov. The clinical investigation was conducted in compliance with the European Directive 93/42/EEC on medical devices [
27] with the Swiss Ordinance on clinical trials of therapeutic products [
28] and with international standards ISO 14155:2011 [
29]. In the absence of standards applicable to our cuffless approach, we evaluated the performance of the app using the standards of the ISO 81060-2:2018 norm [
30].
2.3 Algorithm parameters training
The parameters
\(\widehat{\theta }\) of the non-linear model applied to the set of morphological features extracted by the algorithm were trained using data acquired by a previously published study [
32], minimizing in the least-square sense the error between BP
inv changes (∆BP
inv) and BP
PPG changes (∆BP
PPG). To that end, significant changes in BP in the invasive reference data were selected and compared to their corresponding PPG-derived BP changes. By concatenating all selected ∆BP
inv changes and their corresponding ∆BP
PPG values for all 40 patients in vectors
\({\varvec{U}}\) and
\({\varvec{V}}\) respectively, the parameters
\(\widehat{\theta }\) of the model were optimized in the least-square sense, i.e., by solving
\(\widehat{\theta }={\underset{\theta }{\mathrm{argmin}} \Vert {\varvec{U}}-{\varvec{V}}\Vert }^{2}\). The thus trained model was then applied, with no further adaptation, to the smartphone-derived PPG data in the present study.
2.4 Statistical analysis
The main part of our study focused on assessing BP changes (trending ability) rather than estimating absolute BP values. Hence BP changes between all possible pairs of recordings were computed for each patient and no calibration was needed (Fig.
2). A BP change between two recordings
i and
j was obtained as: ∆BP(
i,
j) = BP(
j) − BP(
i) and done both for BP
inv and BP
PPG, resulting in a list of {∆BP
inv, ∆BP
PPG} data pairs for analysis.
To assess the blood pressure trending ability of OptiBP, we used the four-quadrant (4Q) plot method conjointly with polar plots as proposed by Critchley et al. [
33,
34]. In the 4Q-method, the upper-right and down-left quadrants contain all {∆BP
inv, ∆BP
PPG} pairs showing a concordant direction of change. Hence, the derived concordance rate (CR) represents the percentage of data points in which ∆BP
PPG and ∆BP
inv change in the same direction. Although the 4Q-method is a good mean to assess trending ability it does not allow to realize the magnitude of changes between {∆BP
inv, ∆BP
PPG} pairs. To that end, Critchley suggested to transpose the Cartesian coordinate of the 4Q plots to polar coordinates in so-called polar plots, which enable a quantitative assessments of trending ability. As suggested by the author, we assessed the angular concordance rate at ± 30°, with upper radial limits of ± 5° (mean polar angle) as acceptance limits. To exclude non-significant changes, central exclusion zone of 15% was used for 4Q analysis [
33].
The second part of our analysis aimed to pass a clinical judgement on the agreement between BP
inv and BP
PPG. To this end, we used and adapted Saugel et al. [
35] BP error-grid analysis which defined five risk zones for a BP measurement method based on twenty-five international specialists in anesthesiology and intensive care medicine. Note that this error-grid was first stratified for critical care and perioperative purpose, hence DBP was deliberately excluded due to its minor role as an isolated value in this setting. Saugel defined these five risk zones (A: no risk to E: dangerous risk) as follow: (A) No risk (i.e., no difference in clinical action between the reference and test method), (B) Low risk (i.e., test method values that deviate from the reference but would probably lead to benign or no treatment), (C) Moderate risk (i.e., test method values that deviate from the reference and would possibly lead to unnecessary or missed treatment with moderate non-life-threatening consequences for the patient), (D) Significant risk (i.e., test method values that deviate from the reference and would lead to unnecessary or missed treatment with severe non-life-threatening consequences for the patient), (E) Dangerous risk (i.e., test method values that deviate from the reference and would lead to unnecessary or missed treatment with life-threatening consequences for the patient). Note that this methodology is based on comparison between absolute BP values and in absence of calibration in our setting, we had to transform them into absolute values by calibrating (i.e., adding an appropriate offset) BP
PPG by the average of all BP
inv values. By doing so, we artificially find good agreement between BP
PPG and BP
inv values for patients were there is low BP variability during the measurements. For this reason, all the measurements of patients for whom no significant {∆BP
inv, ∆ BP
PPG} pair was found were entirely rejected for this part of the analysis, thus leaving us only with patients with significant BP variability, thereby providing a more realistic evaluation of the performance of our method.
The last part of our analysis aimed at assessing the ability of OptiBP to accurately estimate BP. Each calibrated BP
PPG value was compared to its corresponding BP
inv value in terms of accuracy (bias) and precision of agreement (SD of the differences) based on the ISO 81060-2:2018 norm [
30]. Due to the absence of an applicable norm for continuous BP measurement devices, the latter was used as a point of comparison. When using invasive continuous data as BP reference, our analysis takes into account the variability of said reference when evaluating the agreement with the device under test. More specifically, as illustrated in the right-hand side of Fig.
2, the ISO 81060-2:2018 standard details that if the BP of the device under test falls within the ± 1 SD interval around the average value of BP
inv, the error is considered to be zero (zero-zone). In addition to providing the accuracy (bias) and precision of agreement (SD) in mmHg, we also provided them as percentage errors, i.e., with normalization of the difference between BP
inv and BP
PPG by the value of BP
inv.
2.5 Sample size determination
The minimal sample size to detect a change of 5 mmHg with a worst-case expected standard deviation (SD) of the error of 12 mmHg at a 5% significance level with a power of 90% was determined to be 61 (two-sided, one-sample test) [
36]. The value of 5 mmHg was chosen because it is below any physiologically expectable 20% change in MBP, whereas the value of 12 was the upper limit of the 95% confidence interval of the SD obtained in a previous PPG-based study by our group during anesthesia induction [
32]. Expecting possible dropouts due to the use of a smartphone (generally lower signal quality than standard pulse oximeters and risk of inadequate finger positioning), a security margin was taken, and 121 patients were enrolled.