Abstract
Background A smartphone is a promising tool for daily cardiovascular measurement and mental stress monitoring. Photoplethysmography (PPG) and low-cost thermography can be used to create cheap, convenient and mobile systems. However, to achieve robustness, a person has to remain still for several minutes while a measurement is being taken. This is very cumbersome, and limits the usage in applications such producing instant measurements of stress.
Objective We propose to use smartphone-based mobile PPG and thermal imaging to provide a fast binary measure of stress responses to an event using dynamical physiological changes which occur within 20 seconds of the event finishing.
Methods We propose a system that uses a smartphone and its physiological sensors to reliably and continuously measure over a short window of time a person’s blood volume pulse, the time interval between heartbeats (R-R interval) and the 1D thermal signature of the nose tip. 17 healthy participants, involved in a series of stress-inducing mental activities, measured their physiological response to stress in the 20 second-window immediately following each activity. A 10-cm Visual Analogue Scale was used by them to self-report their level of mental stress. As a main labeling strategy, normalized K-means clustering is used to better treat interpersonal differences in ratings. By taking an array of the R-R intervals and thermal directionality as a low-level feature input, we mainly use an artificial neural network to enable the automatic feature learning and the machine learning inference process. To compare the automated inference performance, we also extracted widely used high level features from HRV (e.g., LF/HF ratio) and the thermal signature and input them to a k-nearest neighbor to infer perceived stress levels.
Results First, we tested the physiological measurement reliability. The measured cardiac signals were considered highly reliable (signal goodness probability used, Mean=0.9584, SD=0.0151). The proposed 1D thermal signal processing algorithm effectively minimized the effect of respiratory cycles on detecting the apparent temperature of the nose tip (respiratory signal goodness probability Mean=0.8998 to Mean=0). Second, we tested the 20 seconds instant perceived stress inference performance. The best results were obtained by using automatic feature learning and classification using artificial neural networks rather than using pre-crafted features. The combination of both modalities produced higher accuracy on the binary classification task using 17-fold leave-one-subject-out (LOSO) cross-validation (accuracy: HRV+Thermal: 76.96%; HRV: 60.29%; Thermal: 61.37%). The results are comparable with the state of the art automatic stress recognition methods requiring long term measurements (a minimum of 2 minutes for up to around 80% accuracy from LOSO). Lastly, we explored the impact of different data labeling strategies used in the field on the sensitivity of our inference methods and the need for normalization within individual.
Conclusions Results demonstrate the capability of smartphone biomedical imaging in instant mental stress recognition. Given that this approach does not require long measurements requiring attention and reduced mobility, it is more feasible for mobile mental healthcare solution in the wild.