Machine Learning for Stress Detection based on Wearable Sensor Data

This thesis investigates automatic stress recognition from non-invasive, wearable physiological signals using two complementary paradigms: a feature–driven (FD) pipeline based on handcrafted descriptors and classical machine learning, and a data–driven (DD) one-dimensional convolutional neural network (1D-CNN) trained end-to-end. Multi-modal data include electrodermal activity, photoplethysmography, accelerometry, and skin temperature. A unified pre-processing and windowing scheme is adopted; the FD branch extracts time-, frequency-, and non-linear features, followed by model selection among tree ensembles, linear models, and kernels. The DD branch employs a compact CNN architecture with modality-specific streams and late fusion. Robustness is assessed through an in-domain evaluation and a comprehensive cross-dataset protocol spanning four corpora with heterogeneous elicitation protocols and sensors (WESAD, CAMPANELLA, VerBIO, AffectiveROAD). We analyze two operational settings: zero-shot cross-test (train on a source dataset and test on a disjoint target without adaptation) and transfer learning (fine-tuning on limited target data). Macro-F1 is the primary metric due to label imbalance and deployment relevance. Empirically, the FD pipeline offers stronger zero-shot generalisation: considering the best off-diagonal source for each target, FD outperforms DD on all targets (e.g., WESAD: 0.7192 vs. 0.5370; CAMPANELLA: 0.7675 vs. 0.5875; VerBIO: 0.5852 vs. 0.4941; AffectiveROAD: 0.6678 vs. 0.2787). When light adaptation is allowed, the picture reverses: fine-tuning markedly boosts the CNN, surpassing the FD cross-test maxima on three of four targets (up to 0.8030 on WESAD; 0.7713 on CAMPANELLA; 0.6458 on VerBIO), while AffectiveROAD remains marginally in favor of FD (0.6678 vs. 0.6550). In-domain baselines confirm the complementarity of the two families, with CNNs excelling on laboratory-style datasets and FD models remaining competitive where structured, high-SNR descriptors suffice. Taken together, these results suggest a practical deployment guideline: in no-label target scenarios, FD is the safer choice; when even modest target supervision is feasible, adapted DD models become preferable. The observed off-diagonal performance gap is traced to cross-corpus divergences in labeling schemes, protocol stressors, device stacks, and temporal priors. Building on these findings, the thesis outlines a roadmap to bridge the gap—self-supervised pretraining on heterogeneous wearables, domain generalisation without target labels, test-time/source-free adaptation, personalised/federated updates for subject idiosyncrasies, and probabilistic label fusion for ecological validity—aimed at reliable stress sensing beyond controlled settings.