*4.2. Mitigating the E*ff*ects of External Vibrations and Motion Artifacts*

In point-of-care settings, sensing systems for obtaining arterial pulse waveforms often encounter external vibration or motion related artifacts (e.g., battlefield settings or civilian patient transport). These artifacts can greatly impact the quality of the waveforms that are measured, and result in errors in the computation of clinically relevant information (e.g., CRM). External vibrations from transport vehicles during en route care, for example, can be quite large (e.g., on an ambulance or helicopter). Motion artifacts will always be present in the measured signals, unless the patient is unconscious, and may result from either whole-body movements or even more subtle sources such as respiration, talking, or coughing. There are many approaches for mitigating the effects of vibration and motion artifacts in arterial waveform measurements, but the most common techniques involve: (1) improving the signal quality at the source as much as possible, (2) providing auxiliary sensors to detect and cancel motion artifacts from the measured waveforms, and (3) quantifying signal quality on a beat-by-beat basis to facilitate rejection of lower quality waveforms from the subsequent data analysis. Figure 6 shows an example of BP and PPG signals (green, red, and infrared wavelength) captured from a representative subject using a wearable watch technology [101] (a) while at rest (b) and following vigorous exercise with example motion artifacts (c).

**Figure 6.** Example PPG signals taken together with electrocardiogram (ECG) and BP measurements as a reference for comparison. (**a**) The signals were obtained by the Georgia Tech SeismoWatch hardware as described in Ganti, et al. [102]. The BP waveforms shown for comparison were obtained with the ccNexfin volume-clamping based finger cuff BP system (Edwards Lifesciences). (**b**) Signals measured from a subject at rest. Note that the PPG waveforms closely resemble the BP waveforms in shape, with the red and IR (PPG<sup>r</sup> and PPG<sup>i</sup> , respectively) containing many of the same characteristics expected in an arterial pulse waveform, while green (PPGg) appears to be a smoothed version of the BP waveform. (**c**) Signals measured from the same subject following heavy exercise with motion artifacts corrupting the waveforms. The red and IR signals are corrupted heavily while the green PPG signal quality remains high.

#### 4.2.1. Improving the Signal Quality at the Source

For reflectance-mode PPG signals, signal quality is optimized at the source against motion artifacts through the use of green wavelengths rather than red or infrared (IR) [79,102,103]; green penetrates less deeply into the skin, and thus is less attenuated through the forward and backward path through the tissue. Providing non-zero contact pressure between the PPG sensor and the skin can also increase the amplitude of the measured waveforms [53,104]. Specifically, the PPG amplitude is maximized when the contact pressure is equal to the mean arterial pressure (i.e., the transmural pressure is zero). Thus, to reduce the impact of motion artifacts, green wavelengths can be employed for PPG detection, and a non-zero contact pressure can be applied between the sensor and the skin to optimize signal level. The waveforms shown in Figure 6c visually demonstrate this relationship between wavelength and resultant PPG signal quality during motion artifacts. While the red and IR PPG signals are quite heavily affected by the motion artifacts, the green PPG signal quality remains high. Nevertheless, note that many of the key waveform features captured by the red and IR PPG are missing in the green PPG signal due to the fact that the green signal captures primarily the superficial cutaneous vasculature while red and IR penetrate deeper into the skin.

√ For tonometry-based arterial pulse waveform measurements, optimizing signal quality at the source fundamentally requires robust coupling between the superficial artery and the sensor. Tonometry requires a backing force such that the sensor remains consistently in contact with the arterial wall throughout the measurement duration. Thus, a strap is typically used to provide such backing force, for example for radial artery tonometry, and the tightness of the strap must be optimized to be high enough such that the sensor remains in contact with the artery but not high enough to occlude the artery [105]. To reduce the variability due to sensor placement, arrayed sensors are also often used for tonometry based recordings [106]. The sensing system can thus be placed over the palmar aspect of the wrist near the radius bone, and software based approaches can be used to find the waveform with the highest signal quality from the array of sensors.

Techniques for optimizing wearable ultrasound array based arterial pulse waveforms are not yet well understood since the measurement modality is relatively new. However, from an intuitive perspective it is likely that the ability to accurately place the ultrasound array in the proximity of the artery from which measurements will be taken (e.g., the carotid artery), and the coupling between the sensor array and the skin (likely requiring acoustic matching such as ultrasound gel), will play important roles in ensuring high quality waveforms are obtained.

For cardiogenic vibration signals, there are also several aspects that must be considered to optimize signal quality at the source. First, the sensing system should use a sensor with sufficiently low noise floor to capture the micro-vibrations. In the case of SCG signals for example, only accelerometers with input-referred noise of 50 µgrms/ √ Hz or lower should be used. The standard accelerometers deployed on wearable sensing systems and smartphones for inertial measurement have much higher noise than this, with values typically in the 150–300 µgrms/ √ Hz range. Second, leveraging the information from all three axes of the SCG signal, or even including rotational components (i.e., gyrocardiography) as captured with a gyroscope, has been demonstrated to yield greater information than the dorso-ventral axis alone [107]. Third, and perhaps most importantly, the sensor should be rigidly adhered to the body such that movement of the person wearing the sensor does not lead to detachment or other major mechanical disturbances.

#### 4.2.2. Providing Auxiliary Sensors to Detect and Cancel Motion Artifacts

A commonly-used technique for reducing the impact of motion artifacts on PPG signals is the inclusion of an auxiliary accelerometer to detect and provide digital subtraction of motion artifacts [108–110]. The captured acceleration signal provides a noise reference that can be used via adaptive noise cancellation or other signal processing approaches to remove such artifacts from the PPG signal. An alternative approach to reducing motion artifact influence on wearable cardio-mechanical signals leverages auxiliary sensing to capture other signals of cardiovascular origin, namely the electrocardiogram (ECG) [111,112]. Subsequently, rather than removing motion artifacts, the signal strength itself can be bolstered. While the authors are not aware of such auxiliary sensor-based methods for increasing robustness to motion artifacts in tonometry and ultrasound-based arterial waveform capture modalities, intuitively such methods should be directly applicable to these modalities as well. The fundamental approach of either providing a noise reference for noise cancellation or a timing reference for ensemble averaging or otherwise strengthening the signal characteristics are valid for these modalities similarly as for PPG signals.

For cardiogenic vibration waveforms, several approaches have been demonstrated in the existing literature for detecting and cancelling artifacts due to motion or external vibration. Auxiliary sensors for detecting or cancelling motion artifacts from BCG signals include foot electromyogram (EMG) sensing to determine periods of elevated motion as well as external geophone based recordings of floor vibrations for subsequent cancellation [113,114]. Furthermore, signal enhancement using concurrent ECG signals for ensemble averaging, synchronized moving averaging, and otherwise beat segmentation is standard practice.

#### 4.2.3. Quantifying Signal Quality for Rejecting Lower Quality Waveforms

A third approach that can be leveraged to mitigate the effects of external vibrations and motion artifacts on arterial pulse waveforms is the automatic quantification of signal quality on a beat-by-beat basis. Such signal quality assessment is an important tool towards quantifying when the waveforms should be inputted to subsequent machine learning steps (e.g., CRM computation) or, alternatively, when waveform segments should be rejected. Signal quality indices have thus been developed for PPG and cardio-mechanical signals, and have been validated in recent literature [115–117]. The challenge in such algorithms is that both the signal (of cardiac origin) and the noise are non-stationary, and there is substantial variability in signal shape across subjects and also sensor locations. Thus, conventional approaches such as matching the morphology of measured PPG

(or tonometry, ultrasound, cardiogenic vibration signals, etc.) to previous recordings or a database of recordings is not an appropriate technique. Rather, waveform matching must be accomplished using techniques such as dynamic time warping (DTW) [118], which allow for stretching of each beat against the templates with which the beat is compared. DTW-based approaches have demonstrated promise for arterial pulse signals [116]. The establishment of such automated techniques for signal quality assessment—as compared to manual annotation which has been employed in many studies in the existing literature—will represent an important step towards facilitating translation of these sensing approaches to point-of-care settings. Note that, whenever possible, techniques for improving signal quality should be employed rather than techniques for only assessing signal quality. However, in practical settings, many sources of artifacts, noise, and interference corrupting physiological measurements cannot be completely attenuated by signal capture optimization, nor can they be completely removed by auxiliary sensors and associated noise reduction algorithms; thus, the ability to detect and remove low quality events is a key element in delivering robust and reliable CRM outputs to caregivers for subsequent clinical decision making.

#### *4.3. Eliminating the Need for Baseline Measures*/*Calibration*

The use of wearable sensors for CRM-based hypovolemia assessment in field settings may not allow for baseline data to be obtained; for example, if one envisions a person injured in a major car accident, an emergency medical technician (EMT) may simply apply a wearable patch or system to the person when arriving on the scene after exsanguination has begun. Accordingly, algorithms for quantifying compensatory reserve based on machine learning should be globalized rather than designed in a patient-specific manner (see Figure 3). Features leveraged by the algorithm should thus be based on relative measures (e.g., timing intervals, variability measures, etc.) rather than absolute measures (e.g., absolute amplitude of the signal). Moreover, machine learning algorithms should be trained using leave-one-subject-out cross-validation (LOSO-CV) rather than *n*-fold CV, with at least one subject deliberately left out of the training set such that the algorithm focuses on global trends in the sensed waveforms. Finally, since sensor placement can impact the shape of waveforms measured for many reflective PPG [79], tonometry, ultrasound-based blood pressure, and cardiogenic vibration signals [117], such placement-dependent changes should be thoroughly quantified, and methods for harnessing underlying dynamics should be leveraged as compared to features that require manual annotation [118].

#### *4.4. Real-Time Measurements and Processing for Display*

An important consideration is how to display the resultant information derived from the arterial waveforms to the physician or caregiver. One option is to provide a dashboard type display with perhaps a single numerical value indicating the compensatory status of the person (i.e., a CRM value). Another option might be a red, yellow, or green indicator to provide information regarding the clinical decisions to be made during triage (Figure 3). An exciting opportunity exists in the pairing of the volume status information delivered through the automatic analysis of the arterial pulse waveforms with autonomous critical care systems for combat casualty care. Scientists in the academic and commercial domain are conducting research designed to develop systems and methods for providing fluids autonomously to combat casualties based on physiological data [119–122]; providing more in-depth measurements of volume status beyond traditional vital signs to such systems may yield improved results in managing fluid for hemorrhaging patients or casualties. As different applications and use cases emerge, it will be important to determine what processing will be applied at what stage in the system. For example, in one implementation the signals may be wirelessly transmitted from the wearable sensing system to a local smartphone, tablet, laptop, or other dedicated receiver, at which point algorithms may be implemented on that receiver device to output a CRM to be displayed to the caregiver. Another implementation that is possible is to incorporate the CRM machine learning algorithm into the wearable hardware itself (i.e., computing on the edge), in which case the CRM value

itself may be transmitted wirelessly or a readout may be provided on the hardware itself. Regardless of where in the signal chain the processing is implemented, it will be necessary to consider security and privacy concerns of the patient, as well as power consumption and associated battery life on the wearable hardware itself.

#### *4.5. Electronic Documentation in the Prehospital Setting*

The ability to collect and analyze large quantities of data from trauma patients, particularly in austere prehospital settings such as the battlefield, hinder the potential for understanding and improving clinical process and performance [123]. In situations where battery life must be extended for as long as possible, or when wireless transmission is otherwise not feasible, data storage locally on the sensing system may be desired and implemented using micro secure digital (microSD) cards or other non-volatile memory on board the system. The advantage to such local storage of all physiological waveforms is that a detailed record can be kept of the data for subsequent analysis and/or evaluation of the treatment approaches employed. Data extracted from all patients could then be used to retrospectively determine which approaches were most successful, and care can then be optimized accordingly with this evidence. In some instances, the amount of data being stored may be quite large, and may necessitate compressed sensing approaches prior to digitization [124,125]. However, in most cases—since physiological signals such as the PPG are typically of low bandwidth (<100 Hz)—direct digitization and storage of data are feasible for many weeks of continuous recording.
