Pulse Detection with a Single Accelerometer placed at the Carotid Artery: Performance in a Real-Life Diagnostic Test during Acute Hypotension Jens Muehlsteff, Kiran Dellimore, Member IEEE, Vincent Aarts, René Derkx, Christiane Peiker, Christian Meyer
Abstract— Pulse detection via palpation is a basic and essential procedure in daily medical practice. We have been investigating the performance of a single accelerometer placed above the carotid artery, which is one of the recommended locations for manual palpation. A low-cost sensor attached by an adhesive measures accelerations due to carotid dilatations and whole body vibrations. A real-time demonstrator has been developed to classify 10 second- windows in “Pulse”, “Motion” and “No Pulse” and to infer pulse rate. Data were obtained during a scheduled head-up tilt table test (HUTT). Our results show for a subgroup of 10 patients with acute hypotension a wide spread of “good” signal coverage ranging from as low as 37% up to 100%. Key factors compromising the performance in HUTT are motion artifacts, arrhythmias, sensor placement and sensor-skin coupling. In conclusion, pulse detection with a single accelerometer is sufficiently accurate, if good signal coverage can be achieved.
I. INTRODUCTION Pulse detection is an important, potentially life-saving and therefore an often performed test in daily medical routine, for which many different sensing technologies exist [1]. A pulse results from the successful life-sustaining ejection of blood by the heart, which can be observed by a sensing means at a central or peripheral location from the body. In fact, the electrocardiogram (ECG) solely does not provide sufficient information, since its presence doesn’t guarantee that a blood volume is pushed into the arterial system. The most basic approach to assess pulse presence is by manual palpation. An observer places a finger at the carotid artery and feels the pulse of the patient. It has been shown that this method is error prone [2] – e.g. the observer might feel his own pulse – as well as the procedure takes too long, which is in particular an issue during cardiopulmonary resuscitation (CPR). Another need in reliable pulse detection technologies stems from the interest in comfortable cuffless blood pressure (BP) measurements using surrogate measures such as pulse transit time or pulse arrival time to infer BP. Here, at least one pulse signal needs to be detected e.g. when the pressure pulse arrives at a particular body location. These BP inference techniques face challenges e.g. due to poor pulse strengths in hemodynamically compromised patients at low blood pressures or due to peripheral shut down, with the Jens Muehlsteff, Kiran Dellimore, Rene Derkx and Vincent Aarts are with Philips Research Europe, Eindhoven (The Netherlands), e-mail:
[email protected]. Christiane Peiker and Christian Meyer are with UKE Eppendorf University Hospital Hamburg, Division of Cardiology, Pneumology, and Angiology, e-mail:
[email protected].
978-1-4244-9270-1/15/$31.00 ©2015 IEEE
latter being a well-known issue for photo-plethysmography (PPG). A peripheral site is also affected by vasomotion, which leads to frequent recalibrations and compromised accuracy. Therefore, a central site for pulse detection is preferred in order to minimize these issues [3]. Current pulse detection technologies which are established in clinical practice include PPG, sphygmomanometry, ultrasonography and bioelectricalimpedance. All have shortcomings in terms of accuracy, cost, size and ease of application to the clinical workflow. Therefore, there is a need for a simple, low cost and reliable sensor to objectify pulse detection either for spot checks or for continuous monitoring applications preferably from a central site. Today, very sensitive accelerometer (ACC) sensors are commercially available, which make the detection of pulses easily possible. Previously, we demonstrated the basic feasibility of pulse strength assessment with a single ACC at the carotid artery during a head-up tilt table test (HUTT) hemodynamic challenge, which is used for the diagnosis of neurally mediated syncope (NMS) [4]. The objective of the current work is to extend our previous findings investigating the ability to track pulse presence and pulse rate (PR) in a larger patient cohort. We focus on i) assessment of the susceptibility of the approach to artefacts in a real-life diagnostic test and ii) objectification of the performance of a developed pulse detection algorithm and signal classifier during HUTT trained on data from healthy subjects and intensive care patients. II. METHODS TO EVALUATE ACCELEROMETERBASED PALPATION AT THE CAROTID ARTERY A. Carotid artery for pulse detection The carotid location is a prominent site to assess the presence and the strength of a patient’s pulse, because of its central location, its large diameter, and its ease of access. According to Basic Life Support guidelines, pulse presence can be palpated at the carotid at systolic blood pressure (SBP) of at least 60 mmHg [5]. A major challenge in manual palpation is that pulse detection takes too long (often 25s or more [6-8]) and is very unreliable with a reported sensitivity of 90% and specificity of 55% [2]. The carotid artery is relatively close to the skin surface and consequently pulsation can often be observed even visually. For palpation the neck of the patient is brought into a specific posture with easier access to the artery site which also defines the measurement position, which is not possible for continuous monitoring purposes. Interestingly the carotid
434
site is used by the body to measure the systemic blood pressure via baro-receptors reporting their stretching, which is needed for short term control of blood pressure [9]. B. Artery dilatation as a source of the ACC signal - Basic feasibility of monitoring BP changes using ACC Mathematically the expected acceleration signal perpendicular to the skin, Accz, due to carotid dilatation of an idealized circular artery cross-section with changing radius can be described as the second derivative of the area of the artery, which is given by:
𝐴𝑐𝑐𝑧 =
1 𝑑2 𝐴(𝑝𝑇 ) 4𝜋 𝑑𝑡 2
unexplained history of syncope. The study was approved by the local clinical ethical committee as well as by a Philips internal ethical review board and conforms to the principles outlined in the Declaration of Helsinki. All participants provided written consent to participate in the study. The HUTT was conducted according to the guidelines of the European Society of Cardiology and consisted of 3 phases: P1) an initial supine period at rest, P2) a passive standing exercise at a 70° angle, and P3) a supine period at rest. If syncope failed to manifest after a period of 20 minutes during the second phase of the HUTT then nitro-glycerin was administered sublingually to trigger a hemodynamic response.
(1)
The artery area, A, is a function of the transmural pressure p T due to the vessel compliance [11]. The transmural pressure pT is the pressure difference between the external pressure and blood pressure. The observed signal is proportional to the square of the heart rate and is a complex function of the pressure wave and the arterial compliance encoded in the pressure dependence of the expression A(pT). In a simulation study complemented by a first basic experimental test we found that tracking of BP changes during HUTT is in principle feasible [4] as shown in Figure 1. Related research using ACC for pulse and BP tracking has been reported e.g., by Theodor et al. using an implanted sensor [10].
Fig. 2: (left) Photograph of the SENSATRON system attached to a subject (a detailed description of the system can be found in [12]); (right) Location of the accelerometer sensor at the left or right common carotid.
During the HUTT, acceleration signals from the dilatation of the common carotid (Figure 2, right) were measured with a 3-axis ACC sensor (sampled at 125 Hz) attached to a multiparameter, battery operated device (Figure 2 left). One-lead ECG (sampled at 250 Hz) was also acquired synchronously. A complete description of the “SENSATRON” device can be found in [12]. BP was continuously monitored using a Taskforce – Monitor (Graz, Austria) [13]. The characteristics of the patient group with positive result (i.e., fainting) used in this study are summarized below in Table I. TABLE I SUMMARY OF HUTT TEST PATIENT CHARACTERISTICS – POSITIVE OUTCOME Characteristic Value Number of patients
10
Female (%)
3 (30)
Age range (years)
However, in practice, Accz signals are superimposed by other accelerations due to whole-body motion (ballistocardiographic vibrations), localized motion (e.g. head rotation, swallowing, speaking), or sensor tilts in the earth’s gravitational field. Therefore, in general reliable BP tracking using amplitudes is challenging. C. Experimental setup for the head-up tilt table tests Experimental data were gathered during scheduled headup tilt table tests (HUTT) involving 27 patients with an
28-70
Measurements
Fig. 1: left: dependence of the simulated maximum positive acceleration perpendicular to skin surface divided by the squared heart rate versus SBP, (Parameter: pulse pressure PP=[20…40] mmHg and HR = [60…120] bpm), right: observed normalized acceleration signal during HUTT with impending syncope SBP = [60 …140] mmHg.
ECG, ACC, SBP, DBP
ACC = accelerometer, DBP = diastolic blood pressure, ECG = electrocardiogram, SBP = systolic blood pressure
III. RESULTS A. Real-time accelerometer – based pulse detection Figure 3 shows the basic block diagram of the algorithm of the ACC-based pulse detection demonstrator. The algorithm processes as input a 3-axis ACC signal sampled at 125 Hz.
Fig. 3: Block diagram of the algorithm
435
With a proprietary algorithm the ACC signals are transformed into a waveform that looks similar to a photoplethysmogram. An ‘activity level’ is calculated based on the variance of the ACC signal. The reconstructed pulse with ECG as reference signal is pictured in Figure 4 over a 1 minute period in which SBP increases by about 20 mmHg. Obviously the reconstructed pulse signal also contains artefacts, which are automatically detected by an implemented classifier.
Fig. 4: Upper diagram: Reference ECG and SBP changes during 1 min of a HUTT with about 20 mmHg SBP increase; lower diagram: Sequence of the reconstructed pulse signal including artefacts, which are detected in this example by using a threshold of the “activity level”
ECG as reference in all three phases of the HUTT. The middle diagram shows the inferred raw PR from the ACC (blue line) and the ECG-based heart rate (HR) reference (red line). Despite short periods of strong motion, the pulse signal can be reliably detected during the whole HUTT and is identified as “Pulse OK”. Almost no head turns, speaking or movements were observed.
Fig. 6: Patient #2: upper diagram: Spectrogram of the reconstructed pulse, middle: PR from ACC and HR from ECG; lower diagram: Output of the classifier.
Beats and activity level are fed into a selectable either Linear or Support Vector Machine (SVM) classifier, which classifies each 10 second period into ‘Pulse OK’, ‘No Pulse’, or ‘Motion’. The classifier was trained previously on data from healthy subjects and intensive care patients before and judges based on features of the peaks and the ‘activity level’ whether the peaks are classified as valid beats. From the assigned valid beats PR is calculated. Figure 5 shows the User Interface with an example result “Pulse OK” when mimicking a pulse spot check similar to a clinical scenario with a healthy subject. The peaks (green ‘o’) of this waveform are found with a peak detector. Fig. 7: Patient #20: upper diagram: Spectrogram of the reconstructed pulse, middle: PR from ACC and HR from ECG; lower diagram: Output of the classifier; This patient shows also a few instances in which “No Pulse” is reported.
Fig. 5: Graphical user interface of an accelerometer-based pulse detection demonstrator; (left) reconstructed pulse; (right) classification.
B. Pulse detection with an accelerometer placed at the carotid Figure 6 shows in the upper panel a spectrogram (10 s window and 90% overlap) for calculation of the reconstructed pulse signal derived from ACC. The patient experiences a NMS at about 2000 s. The maximum power in the pulse signal coincides with the heart rate obtained from
In contrast Figure 7 presents similar plots for a patient with significantly lower signal quality. PR can be obtained accurately only when the subject is at rest. There are many periods during standing and after the patient was tilted back, in which the ACC-pulse signal is unreliable. In particular, during the onset of syncope with an increase in PR followed by a decrease in SBP, the signal quality is severely compromised. The patient speaks, starts countermeasures such as head or leg movements in order to increase venous return, or swallows more often. Also, after fainting when the patient is returned the supine position, involuntary movements are often experienced, which result in an unreliable PR.
436
C. Pulse tracking during HUTT in patients with impending syncope – Linear classifier results Table II summarizes the results of pulse tracking for each patient with positive HUTT e.g., manifested syncope. It lists the coverage of “No motion” for the three HUTT phases. The results vary widely from almost complete coverage > 90 % in all HUTT phases but there are also severe outliers such as patients #4, #6, and #7. For patient #4, the low coverage in P2 and P3 is due to strong arrhythmias which to date the pulse detector hasn’t been trained to handle. Patient #6 has an overall low signal quality indicating issues in sensor placement. Obviously, the best average coverage with 89 % is achieved in P1 and reduces to an average of 83 % in P2. During P3, after NMS, pulse detection is compromised (82 % average coverage) due to NMS recovery. TABLE II COVERAGE AND AVERAGE DIFFERENCE OF PR AND HR FOR EACH HUTT PHASE SEPARATELY Coverage [%] Rate Difference [bpm] Patient Id
P1
P2
P3
P1
P2
P3
2 4 6 7 9 20 21 23 31 32
96 87 72 100 100 100 76 81 100 82
100 65 37 57 99 92 96 95 95 95
100 66 20 85 100 90 83 81 96 95
-4.6 -8.3 -10.5 1.8 0.1 -1.0 -13.2 -1.9 -0.1 0.9
-0.2 -24.8 -11.5 -1.1 1.7 -8.3 -16.7 -2.3 -2.7 -3.8
-0.9 -13.5 -1.9 -0.9 0.1 -25.0 -14.8 -0.5 2.5 -0.1
Average Std Minimum Maximum
89 11 72 100
83 22 37 100
82 24 20 100
-3.7 5.2 -13.2 1.8
-7.0 8.4 -24.8 1.7
-5.5 9.0 -25.0 2.5
the classifier as “No motion”, the presence of a pulse can be reliably detected with sufficient PR accuracy. Our results show, that the performance of an ACC for pulse tracking at the carotid is compromised in a real-life diagnostic procedure such as HUTT including a hemodynamic challenge. A reliable tracking of BP trend as we have shown in [4] under an idealized measurement situation is typically not possible. Detection of short-term hemodynamic changes, which is needed for a continuous monitoring approach, in its current embodiment requires improvement. Better sensor attachment e.g. with additional applied contact pressure, signal fusion approaches, e.g., with a PPG sensor or improved signal processing techniques could be tested in future. However, a low cost ACC placed at the carotid for pulse palpation may still be an option under well-controlled conditions with low body motions as well as good sensor attachment. This is typically the case, when a measurement is supervised by a trained person or a patient is unconscious. In these situations, our approach could outperform the current “gold standard” of manual palpation, since it is relatively simple, easy to apply and objective. One interesting application we currently exploring is for CPR to replace manual palpation during compression pauses. REFERENCES [1] R. Neumar, C. Otto, M. Link, et al. “Part 8: adult advanced
[2] [3]
Coverage is a measure of periods without motion artifacts during HUTT P1: supine, P2: passive standing, P3: supine; Rate difference is the 10 s average of the difference in PR derived from ACC and HR obtained from the ECG
[4]
Regarding PR inference we obtain mixed results of very accurate PR tracking compared to an ECG-based HR reference with an average error less than 1 bmp, but also, differences of > 20 bpm, when the beat detector and/or classifier underperforms. The very large difference for patient #21 with low motion levels and therefore large coverage in P2 is due to severe arrhythmias, which appear persistently and was not accounted for in the current algorithm implementation.
[5]
IV. DISCUSSION
[9]
The strength and quality of the cardiac pulse detected by an ACC attached at the common carotid is influenced by many factors which may limit the ability of a single sensor to reliably continuously track the pulse in a conscious, active patient. Artefacts arise from head rotations, swallowing, whole-body movements and speaking during the HUTT procedure. Other issues include phases of arrhythmias, which have been often observed in this patient population as well. Nevertheless, when a subject was at rest identified by
[6] [7] [8]
[10] [11] [12]
[13]
437
cardiovascular life support: 2010 American Heart Association Guidelines for Cardiopulmonary Resuscitation and Emergency Cardiovascular Care.” Circulation. vol. 122, pp. S729-67, 2010. B. Eberle, W. Dick, T. Schneider, et al., "Checking the carotid pulse check: diagnostic accuracy of first responders in patients with and without a pulse," Resuscitation, vol. 33, pp. 107-116, 1996. J. Sola, M. Proenca, D. Ferrario, et al. “Non-invasive and nonocclusive blood pressure estimation via a chest sensor,” IEEE Trans Biomed Eng. vol. 60, pp. 3505-13, 2013. J. Muehlsteff, K. Dellimore, V. Aarts, et al., “Feasibility of pulse presence and pulse strength assessment during head-up tilt table testing using an accelerometer located at the carotid artery,” Conf Proc IEEE Eng Med Biol Soc. 2014, pp. 894-7. C. Deakin and J. Low. “Accuracy of the advanced trauma life support guidelines for predicting systolic blood pressure using carotid, femoral, and radial pulses: Observational study,” BMJ, vol. 321, pp. 673-674, 2000. S. Brearley, C. Shearman, and M. Simms, “Peripheral pulse palpation: an unreliable physical sign,” Ann R Coll Surg Engl., Vol. 74(3), 1992, pp. 169–171. M. Lundin, J. Wiksten, T. Peräkylä, et al., “Distal pulse palpation: is it reliable?” World J Surg., vol. 23, pp. 252-5, 1999. C. Graham and N. Lewis, “Evaluation of a new method for the carotid pulse check in cardiopulmonary resuscitation,” Resuscitation, vol. 53, pp. 37-40, 2002. H. Rau and T. Elbert. “Psychophysiology of arterial baroreceptors and the etiology of hypertension,” Biol Psych., vol. 57, pp. 179-201, 2001. M. Theodor, J. Fiala, D. Ruh, et al. “Implantable accelerometer system for the determination of blood pressure using reflected wave transit time,” Sensor Actuat A-Phys, vol. 206, pp. 151-8, 2014. G. Drzewiecki, J. Pilla. “Noninvasive measurement of the human brachial artery pressure-area relation in collapse and hypertension,” Ann Biomed Eng. vol. 26, pp.965-74, 1998. J. Muehlsteff, P. Carvalho, J. Henriques, et. al., “Cardiac Status Assessment with a Multi-Signal Device for Improved Home-based Congestive Heart Failure Management,” Conf Proc IEEE Eng Med Biol Soc. 2011, pp. 876-9. Taskforce Monitor, http://www.cnsystems.at/products/task-forcemonitor, Accessed online: March 19th, 2014.
