• Keine Ergebnisse gefunden

Compliance and Calibration

In multi-modal EE estimation, advantages of physiological predictors are combined with objective observation of motion (e.g. acceleration, sec. 5.3). However, compliance

5Indeed, handling misclassification of unknown activities and transitions in-between certain activities is an active branch of HAR research[193]

5.4 Compliance and Calibration

and calibration issues remain, which was shown in [272]. It was highlighted that valid measurement of vital signs and reliable measurement of motion depends on prior calibration and appropriate use of the WBSs. This involves the correct attachment and application of a WBS, as well as the pre-execution of calibration or validation steps.

Here, a summary of the signal disturbances and countermeasures presented in[272] is given because they were reused for the experiment presented in sec. 5.5.

5.4.1 Signals and Disturbances

To obtain reliable sensor readings, a variety of disturbances have to be considered.

Taking the functional principle of the sensors into consideration, the necessary steps to ensure accuracy and precision can be determined. In the subsequent, an overview of common disturbances affecting the BG-V4.2 is outlined.

5.4.1.1 Electrocardiogram

Typical disturbances affecting the ECG include electrical interference, poor electrode contact (and skin conductance) as well as the electrode’s positioning[90].

Examples concerning electrical interference are found with the line noise (AC noise 50 or 60 Hz) or high-frequency noise from the electrical activity of the skeletal muscles (EMG noise). In order to remove such electrical interference, analog filters can be used[90]. In BG-V4.2 and BG-V5, an analog active low pass filter of 4th order is used (cut-off frequency 33.86 Hz). To remove the baseline wander, an additional high pass filter is applied (cut-off frequency 1.56 Hz).

Dry electrodes or insufficient skin contact are other common reasons for measurement errors. Regarding textile electrodes, which are commonly used in chest straps, proper moistening is required. Additionally, a minimal tension of the chest strap is needed to ensure good skin contact. Setting the chest strap’s tension is also relevant in terms of motion artifacts, which often occur during vigorous physical activity[173]. Obviously, the positioning of the electrodes or their contact cannot be corrected by the device itself. Therefore, these prerequisites must be checked by the experimenter or the wearer (user) before an experiment is carried out.

5.4.1.2 Accelerometer

Considering motion sensors, the initial calibration by the manufacturer guarantees precise measurements. Also, random and systematic errors are relatively small. For instance, the random observational error of theLIS331HH6accelerometer integrated into the BG-V4.2 is in the order of magnitude of 10−3g. Likewise, systematic obser-vational errors are bounded to approximately 10−4g (e.g. caused by temperature

6STMicroelectronics, LIS331HHhttps://www.st.com/resource/en/datasheet/lis331hh.pdf

fluctuation). However, these systematic observational errors are typically dominated by an orientation offset. This offset results from a misalignment between the sensor’s frame and the body’s frame of the wearer. The latter depends on how and where the chest strap was applied to the wearer’s body[239, 246].

5.4.2 Calibration and Noise Detection

The impacts of noise or other disturbances are widely addressed in recent research.

Likewise, methods to detect or minimize the effects of disturbances are proposed:

Concerning HR estimation, the effect of noise in the ECG-signal was investigated by Friesen et al.[90]. Within their work, they demonstrated the vulnerability of common algorithms used to estimate HR, in case the underlying ECG-signal is affected by different types of noise. Furthermore, concerning mobile HR acquisition, Nikolic-Popovic et al.[173]explain the effect of motion artifacts for HR variability estimation. In order to detect noisy signals, and thus to prevent false alarms or misleading information, several methods are known to detect the quality of the ECG-signal[195].

Likewise, the calibration of accelerometers used for physical activity monitoring is important. In this respect, Wang et al.[246]examined the impact of the orientation error of an accelerometer. They showed that an orientation error greater than 3°

adversely affects the PA estimation. Similarly, Alinia et al.[10]examine a scenario in which the position of a WBS was interchanged. They demonstrated that without knowing the real position of the WBS, accurate estimation of physical activity was impossible. This was again confirmed on the example of HAR by Yurtman et al.[265], who found deviation in accuracy up to 18.8 %. An overview of different approaches to calibrate accelerometers is given by Won et al.[259].

Taking these examples, it can be argued that without a precedent check-up and calibration of a WBS, reliable data is unobtainable. As exemplary pictured above, this applies to accelerometers and ECG recordings. Nevertheless, other sensory elements can be affected as well (e.g. respiration sensor[269]).

5.4.2.1 Heart Rate Validation

Various methods exist to assess the quality of an ECG recording[73]. However, these methods are often developed for clinical investigations and are not optimized for efficiency. Aiming towards an implementation for a WBS, which offers limited resources, less complex solutions are preferable. In addition, the use-cases associated with WBSs application typically require valid HR estimation only. A medical examination or quality criterion is therefore not necessary.

In order to obtain the ECG’s quality in real-time, the solution presented in[138]was adopted and tuned. In[138], multiple weak metrics are combined into one strong predictor applicable for clinical usage. Here, this approach was partly reused, but restricted to use only one of the weak predictors, which is the signal’s kurtosis.

5.4 Compliance and Calibration

Because kurtosis is a measure of the probability distribution’sspikiness, it is well suited to distinguish a valid ECG from white noise. Yet, it is not applicable to detect scattered spikes, which occur due to motion artifacts (chest strap temporally loses contact with the skin). Therefore, an additional rule was added, which marks the signal as invalid if its range exceeds 75 % of the total measuring range (3072 LSB; 12-Bit resolution). The final quality measure combines the 2 decision rules (eq. 5.13).

QualityECG=K(ECG)5.4 ∧R(ECG)3072 (5.13) Kurtosis: K(X) = 1nPni=1€σxi(Xx¯)Š4

Range: R(X) = max(X)min(X)

To test its feasibility, it is validated against thePhysioNet MIT-BIH Noise Stress Test Database7[96]. Each time series in the data set contains 50 % of noise-free and 50 % of noisy data. As a result, it is found that the quality measure is suitable to detect large disturbances with a signal-to-noise ratio (SNR) lower than 6 dB (Table 5.1). In the presence of almost undisturbed signals (SNR≥18 dB), no disturbances are detected.

Thus, the signal is marked as valid. Correspondingly, the accuracy and the positive predictive value considerably drop. However, given a SNR greater than 6 dB, accuracy is at least 96 %.

It becomes clear that only prominent disturbances are detected with the presented approach. Still, the number of false alarms (due to false negatives) is effectively limited.

Therefore, non-acceptable ECG records can be detected and falsely calculation of the HR can be prevented. Moreover, the metric is a valuable tool for the experimenter and wearer to evaluate the ECG recording in advance to an experiment.

Table 5.1:Accuracy (ACC), positive (PPV), and negative predictive value (NPV) of the ECG-quality prediction, evaluated against PhysioNet MIT-BIH Noise Stress Test Database.

data set: 118 data set: 119

SNR ACC PPV NPV ACC PPV NPV

No noise 95 % 100 % 95 % 100 % 100 % 100 %

−6 dB 96 % 100 % 92 % 100 % 100 % 100 % 0 dB 96 % 100 % 92 % 100 % 100 % 100 % 6 dB 96 % 100 % 92 % 97 % 93 % 100 %

12 dB 93 % 95 % 92 % 64 % 28 % 100 %

18 dB 76 % 60 % 92 % 50 % 0 % 100 %

24 dB 51 % 10 % 92 % 50 % 0 % 100 %

7PhysioNet MIT-BIH Noise Stress Test Database,https://physionet.org/content/nstdb/1.0.0/

5.4.2.2 Acceleration and body posture calibration

Regarding acceleration readings, the sensor’s frame (Figure 5.1) and the user’s body frame need to be aligned. To correct the alignment, the participants are asked to adjust the sensors’ position manually. Next, acceleration data is recorded while the user is holding a reference position. Therefore, the user is asked to keep its back straight while standing against a wall.

Here, the assumption is made that the deviation between the body’s frame and the sensor’s frame is high and depends on how the participant applied the sensor. In contrast, the deviations based on possible variations in the body posture are assumed to be small. Apparently, with this approach, no definite statement can be made for therealalignment. Nevertheless, taking a reference posture to distinguish the users’

body frame can improve inter-individual accuracy. Hence, making it more reasonable to compare data among various participants.

To obtain the rotation matrix between the sensor and the user’s body frame, up to 10 s of sensor readings are recorded during which the participant is asked to keep the reference posture. If the mean absolute deviation is small (≤10−6g), the sensor readings are mapped to a reference vectorvr = [0, 0,−1]g. Therefore, the method proposed by[101]based on Rodrigues’ rotation formula is used. An implementation is found in[178](eq. 5.14).

Lateral

X axis Y axis

Z axis +

Yaw axis

Pitch axis +

-+

-Posterior

Roll axis Anterior

Inferior Superior

Figure 5.1:Orientation of the BG-V4.28.

R(a,b) = (5.14)

I3+vx+v2x·1−(a·b) kv2k v=a×b

vx =

0 −v3 v2 v3 0 −v1

v2 v1 0

R Rotation matrix I Identity matrix a,b acceleration vectors

vx skew-symmetric cross-product