Christelle Godin

Wearable cardiac sensors pave the way for advanced cardiac monitoring applications based on heart rate variability (HRV). In real-life settings, heart rate (HR) measurements are subject to motion artifacts that may lead to frequent data loss (missing samples in the HR signal), especially for commercial devices based on photoplethysmography (PPG). The current study had two main goals: (i) to provide a white-box quality index that estimates the amount of missing samples in any piece of HR signal; and (ii) to quantify the impact of data loss on feature extraction in a PPG-based HR signal. This was done by comparing real-life recordings from commercial sensors featuring both PPG (Empatica E4) and ECG (Zephyr BioHarness 3). After an outlier rejection process, our quality index was used to isolate portions of ECG-based HR signals that could be used as benchmark, to validate the output of Empatica E4 at the signal level and at the feature level. Our results showed high accuracy in estimating the mean HR (median error: 3.2%), poor accuracy for short-term HRV features (e.g., median error: 64% for high-frequency power), and mild accuracy for longer-term HRV features (e.g., median error: 25% for low-frequency power). These levels of errors could be reduced by using our quality index to identify time windows with few or no data loss (median errors: 0.0%, 27%, and 6.4% respectively, when no sample was missing). This quality index should be useful in future work to extract reliable cardiac features in real-life measurements, or to conduct a field validation study on wearable cardiac sensors.

Wearable cardiac sensors pave the way to advanced cardiac monitoring applications based on heart rate variability (HRV). In real-life settings, heart rate (HR) measurements are subject to mo-tion artifacts that can be timely removed from the recordings. This leads to frequent data loss in the HR signal, especially for commercial devices based on photoplethysmography (PPG). The cur-rent study had two main goals: (i) to provide a white-box quality index that estimates the amount of missing samples in any piece of HR signal; and (ii) to quantify the impact of data loss on feature extraction in a PPG-based HR signal. This was done by comparing real-life recordings from com-mercial sensors featuring both PPG (Empatica E4) and ECG (Zephyr BioHarness 3). After an out-lier rejection process, our quality index was used to isolate portions of ECG-based HR signal that could be used as benchmark, to validate the output of Empatica E4 at the signal level and at the feature level. Our results showed high accuracy for estimating the mean HR, poor accuracy for short-term HRV features and moderate accuracy for longer-term HRV features. Levels of error could be substantially reduced by using our quality index to identify time windows with few or no missing data.

Objectives : Physiological sensors appear a promising tool to monitor psychological stress in the daily life. Yet, despite the increasing know-how developed in affective computing, physiology-driven stress detection algorithms still yield unsatisfactory results on the field. This study intends to highlight physiological patterns that are both stress-sensitive and stress-specific, at the multi-subject scale, in several contexts. Approach : A notional framework was developed and empirically validated, which distinguishes psychological stress from two context-related confounding factors: Background and Activity. A physiological database was elaborated from a multi-task experimental protocol in a standardized environment with skin conductance, electrocardiograph and respiration recordings. Stress-induced variations were isolated in standard physiological features, to select the most stress-sensitive ones independently from confounding factors. Main Results : Most standard features were unable to distinguish acute psychological stress from low cognitive activity (e.g. frequency of skin conductance responses, or spectral power in the low and high frequency bands of the interbeat intervals). After sign-rank tests between the stressful and control conditions of two standard tasks involving arithmetic or speech, their effect sizes remained below 0.3. A number of physiological features also showed task-dependency. Yet, promising correlates were highlighted among cardiac features such as mean heart rate and its standard deviation, or the root mean square of its successive differences (effect sizes above 0.3 for both tasks). Significance : The high sensitivity to context observed in standard features is likely to explain the poor validity on the field of physiology-driven models developed in controlled environments. In particular, mental activity should be considered as confounding factor while elaborating these models to increase their accuracy. The small feature set highlighted in this paper and the use of control data should enhance the generalization abilities of any machine learning algorithm; still, further research is needed to evidence new stress-specific physiological responses.

This article addresses the question of passengers’ experience through different transport modes. It presents the main results of a pilot study, for which stress levels experienced by a traveller were assessed and predicted over two long journeys. Accelerometer measures and several physiological signals (electrodermal activity, blood volume pulse and skin temperature) were recorded using a smart wristband while travelling from Grenoble to Bilbao. Based on user’s feedback, three events of high stress and one period of moderate activity with low stress were identified offline. Over these periods, feature extraction and machine learning were performed from the collected sensor data to build a personalized regressive model, with user’s stress levels as output. A smartphone application has been developed on its basis, in order to record and visualize a timely estimated stress level using traveler’s physiological signals. This setting was put on test during another travel from Grenoble to Brussels, where the same user’s stress levels were predicted in real time by the smartphone application. The number of correctly classified stress-less time windows ranged from 92.6% to 100%, depending on participant’s level of activity. By design, this study represents a first step for real-life, ambulatory monitoring of passenger’s stress while travelling.

Quite a few physiological measures have been shown to correlate with psychological stress, and used to design some stress detection algorithms. As most of them are also known to react to non-affective mental states as well, physiology-driven stress models based on these features may lack from specificity. This article tackles this issue by looking for physiological features that are both responsive to affective stress load and minimally sensitive to cognitive activity. A laboratory experiment was designed in such a way that activity-related variations in physiological measures can be isolated during exposure to common tasks. A pressure component was defined and extracted from our data. Its relevance as a predictor of features' ability to distinguish between the stressful and control conditions was shown from descriptive statistics. A feature selection procedure was applied on usual heart rate, heart rate variability and skin conductance-based features, to identify the most pressure-specific ones in a between-subjects perspective. Finally, a simple model built from the selected features succeeded in detecting 87.5% of stress examples among instances of stressless task performance.

A mono-feature fuzzy index that evaluates the stress level from one feature extracted from ECG or GSR is presented. It is build using several measures of the feature recorded when the subject is at rest. The mono-feature fuzzy index can be merged in a multi-feature stress index without any tuning. It can be used to select relevant features and to detect stress. The performance of the stress index is analyzed on a data set made of 160 time periods of time when 20 subjects had to perform stressful tasks and corresponding control tasks. The stress was induced by 4 different tasks. The performances reached are 72% of correctly classified time periods in stress and no stress situations. Interesting conclusions could also be made on the tasks ability to induce stress.

To test the stress detection performance of a wearable sensor system, the signals related to heart activity and electrodermal activity of the Empatica E4 wristband have been compared to stationary electrocardiogram and finger skin conductivity electrodes of high sampling rates during the classical laboratory stress protocol Trier Social Stress Test. The comparison has been done on both signal level and in terms of features for stress detection on a total of seven subjects. The main results point out that the Empatica E4 wristband had a significant loss in terms of detected interbeat intervals, but that time-domain features such as the mean heart rate and standard deviation of the heart rate were still well estimated, with good stress discrimination power. Furthermore, the skin conductivity signals measured at different locations (wrist versus finger) show no visual resemblance and it appears that the signal from the Empatica E4 wristband yielded higher stress discrimination power than the signal measured at the fingers.