Abstract:
As artificial intelligence-enabled wearable devices become increasingly integrated into everyday life, they offer new opportunities to continuously monitor physiological signals and support emotional well-being outside clinical settings. Such devices are already capable of monitoring basic health indicators, such as activity levels and heart rate, yet their ability to reliably infer emotional states and support real-world monitoring of mental well-being remains limited. Early research in physiological emotion recognition demonstrates technical feasibility within lab settings, but several challenges must be addressed before these algorithms can be effectively deployed at scale in real life. Key challenges include data availability, data consistency, human-centred validity, and real-world usability. Large, diverse, high-quality labelled emotion datasets remain scarce, limiting robust model development. Existing datasets are often heterogeneous in sensing modalities, devices, experimental protocols, and annotation methods, which complicates benchmarking and hinders model generalisation across contexts. Additionally, emotion labelling inherently involves subjective interpretation, yet many data collection practices insufficiently incorporate participants’ lived experiences, leading to inconsistent annotations and variable data quality. Finally, even technically capable models must align with user needs, expectations, and deployment realities to deliver meaningful support for mental well-being. With everyday mental well-being as the primary application, this dissertation addresses these challenges by (1) systematically analyzing and benchmarking heterogeneous physiological emotion datasets, (2) understanding participant’s perspectives and developing human-centered approaches for collecting more authentic and reliable emotion-labeled data, (3) introducing new datasets and modeling approaches aimed at improving robustness and generalizability, and (4) examining user perspectives to inform the design of deployable mental well-being technologies. Together, these contributions advance the design of physiological emotion datasets, data-collection methodologies, and modelling strategies aimed at improving ecological validity, robustness, and generalisability in emotion recognition research. This dissertation, therefore, offers a comprehensive framework to support the development of data-driven interventions for emotional well-being in everyday settings. Collectively, the work moves the field closer to a future in which wearable intelligent systems can more reliably interpret emotional experiences and deliver meaningful, context-sensitive support for mental well-being in everyday life.