Stress is an increasingly prevalent mental health condition across the world. In Europe, for example, stress is considered one of the most common health problems, and over USD 300 billion are spent on stress treatments annually. Therefore, monitoring, identification and prevention of stress are of the utmost importance. While most stress monitoring is carried out through self-reporting, there are now several studies on stress detection from physiological signals using Artificial Intelligence algorithms. However, the generalizability of these models is only rarely discussed. The main goal of this work is to provide a monitoring proof-of-concept tool exploring the generalization capabilities of Heart Rate Variability-based machine learning models. To this end, two Machine Learning models are used, Logistic Regression and Random Forest to analyze and classify stress in two datasets differing in terms of protocol, stressors and recording devices. First, the models are evaluated using leave-one-subject-out cross-validation with train and test samples from the same dataset. Next, a cross-dataset validation of the models is performed, that is, leave-one-subject-out models trained on a Multi-modal Dataset for Real-time, Continuous Stress Detection from Physiological Signals dataset and validated using the University of Waterloo stress dataset. While both logistic regression and random forest models achieve good classification results in the independent dataset analysis, the random forest model demonstrates better generalization capabilities with a stable F1 score of 61%. This indicates that the random forest can be used to generalize HRV-based stress detection models, which can lead to better analyses in the mental health and medical research field through training and integrating different models.

Thanks to wearable devices joint with AI algorithms, it is possible to record and analyse physiological parameters such as heart rate variability (HRV) in ambulatory environments. The main downside to such setups is the bad quality of recorded data due to movement, noises, and data losses. These errors may considerably alter HRV analysis and should therefore be addressed beforehand, especially if used for medical diagnosis. One widely used method to handle such problems is interpolation, but this approach does not preserve the time dependence of the signal. In this study, we propose a new method for HRV processing including filtering and iterative data imputation using a Gaussian distribution. The particularity of the method is that many physiological aspects are taken into consideration, such as HRV distribution, RR variability, and normal boundaries, as well as time series characteristics. We study the effect of this method on classification using a random forest classifier (RF) and compare it to other data imputation methods including linear, shape-preserving piecewise cubic Hermite (pchip), and spline interpolation in a case study on stress. Features from reconstructed HRV signals of 67 healthy subjects using all four methods were analysed and separately classified by a random forest algorithm to detect stress against relaxation. The proposed method reached a stable F1 score of 61% even with a high percentage of missing data, whereas other interpolation methods reached approximately 54% F1 score for a low percentage of missing data, and the performance drops to about 44% when the percentage is increased. This suggests that our method gives better results for stress classification, especially on signals with a high percentage of missing data.

Wrist-worn sensors provide continuous physiological time series, but using them to monitor stress requires a label built from three choices: the person's normal reference state against which the signal is contrasted, the temporal width assigned to a self-report, and the direction in which that report is matched to the signal. We quantified how much these choices change the measured association, using a dataset in which the objective context and the perceived stress were recorded independently over the same periods. Twenty-six students wore an Empatica EmbracePlus for several weeks, providing electrodermal activity, blood volume pulse, skin temperature and acceleration, alongside a dated calendar of academic events and ecological momentary assessments. Electrodermal activity was decomposed into tonic and phasic components, pulse-to-pulse intervals were corrected for ectopic and missed beats, and modality-specific quality criteria were applied before any contrast. The same participants completed a controlled session used to check that the chain detects a physiological response under aligned conditions. Contrasts were computed at the participant level as Cliff's delta with bootstrap confidence intervals, and differences between contrasts were tested directly. Examination periods were associated with lower cardiac variability relative to baseline, whereas high perceived-stress reports were not. Each of the three labelling choices substantially changed the measured effect size, by up to a factor of 1.7 on the same recordings. Label definition should therefore be reported, varied, and treated as a parameter of the processing pipeline, on the same footing as filtering or artefact rejection.