Low-cost wrist-worn sensing for scalable, low-burden assessment of mental-health symptoms is a central topic in mobile health. In this paper, we propose MentalCare, a multi-class psychiatric disorder (PD) detection system that integrates ubiquitous multi-source sensors, including Pulse sensor, LMT70, Grove GSR, and Max30102. These sensors capture physiological signals such as photoplethysmography (PPG), skin temperature (SKT), galvanic skin response (GSR), and blood oxygen (OXY). To address physiological heterogeneity among individuals, we design a contrastive disentanglement with a positive transfer framework (CDPT), which consists of a stacked convolutional neural network module (SCNN), a cross-modal attention interaction module (CMAI), a subject-level contrastive disentanglement module (SCD), and an adaptive feature selector module (AFS). In particular, the SCD module employs supervised contrastive learning to promote the separation of heterogeneous features among different subjects in the latent subspace. Furthermore, the proposed AFS module selects features from the abundant source-domain heterogeneous representations that are most similar to the target domain, thereby facilitating feature alignment and enhancing detection performance. We collect a multi-class PD dataset comprising 74 participants diagnosed with depression, bipolar disorder, anxiety, or verified as healthy controls. The experiment result demonstrates that CDPT achieves 83.96% overall accuracy under the leave-one-subject-out cross-validation protocol, significantly outperforming existing state-of-the-art methods. At the system level, we further validate the effectiveness and rationality of the proposed system when deployed on edge devices.

Depression significantly impacts mental health, severely disrupting patients' daily lives. During depressive episodes, individuals may experience symptoms such as excessive guilt, self-harm, and suicidal ideation. Compared to proprietary devices like brain electrode caps, wearable technologies for depression detection have gained attention due to their affordability and portability—enabling real-time monitoring of depressive states. However, challenges such as low-quality data from ubiquitous devices, individual variability, and the complexity of multimodal physiological signal analysis limit model generalizability. To address these issues, we present DepGuard, a novel ubiquitous wearable system for depression assessment based on multimodal physiological signals. DepGuard performs a two-stage detection process: depression recognition and real-time episode monitoring. For depression recognition, we propose an unsupervised domain adaptation method to reduce the domain gap between source and target subjects. For episode monitoring, we employ a few-shot learning strategy to enable personalized modeling. Both approaches enhance cross-subject generalization. Our system achieves 90.75% accuracy in cross-subject depression recognition using 30 unlabeled samples per target subject, and 93.52% accuracy in episode monitoring using 15 labeled samples per class.

Psychiatric disorders represent a growing global public health concern. Current screening methods primarily rely on subjective self-reports, which prone to bias and often impede timely intervention. In this paper, we propose G-PsychSW (Generalizable Psychiatric Screening from Wearable), a continuous and objective mental health monitoring based on multi-channel biosignals, including photoplethysmography (PPG), galvanic skin response (GSR), skin temperature (SKT), and blood oxygen (OXY). To address challenges such as individual variability and limited interpretability, we introduce causal disentanglement to learn subject-invariant representations to enable domain generalization (DG). We further incorporate multi-scale attention convolution (MAConv) and functional graph construction (FGC) to model temporal dynamic and inter-channel interaction, uncovering meaningful interpretability analysis corresponding to psychial biomarkers. Experiments on a customized dataset under leave-one-subject-out cross-validation show that G-PsychSW achieves 8 0. 1 6% accuracy and 8 6. 5 2% F1-score across four-class psychiatric screening, including healthy, depression, anxiety and bipolar disorders, outperforming state-of-the-art DG methods and validating its effectiveness.

Acoustic-based human gesture recognition (HGR) offers diverse applications due to the ubiquity of sensors and touch-free interaction. However, existing machine learning approaches require substantial training data, making the process time-consuming, costly, and labor-intensive. Recent studies have explored cross-modal methods to reduce the need for large training datasets in behavior recognition, but they typically rely on open-source datasets that closely align with the target domain, limiting flexibility and complicating data collection. In this paper, we propose${\sf Img2Acoustic}$, a novel cross-modal acoustic-based HGR approach that leverages models trained on open-source image datasets (i.e., EMNIST, Omniglot) to effectively recognize custom gestures detected via acoustic signals. Our model incorporates a task-aware attention layer (TAAL) and a task-aware local matching layer (TALML), enabling seamless transfer of knowledge from image datasets to acoustic gesture recognition. We implement${\sf Img2Acoustic}$on commercial devices and conduct comprehensive evaluations, demonstrating that our method not only delivers superior accuracy and robustness compared to existing approaches but also eliminates the need for extensive training data collection.

The rising prevalence of mood disorders, including depression and anxiety, underscores the critical state of physical and mental health issues in contemporary society. Automatic emotion recognition technology emerges as a potential tool for monitoring mental health disorders, and offering valuable guidance through human-computer interfaces. However, existing technologies grapple with limitations, such as emotional concealment, potential privacy leakage, and specific device location restrictions. To overcome these limitations, we propose a mobile wearable emotion recognition system called EmoTracer incorporating ubiquitous multi-source sensors to measure physiology signals, and address challenges such as the intricate signal-emotion relationship, sparse data processing, and notable disparities among different subjects. We implement a real-time prototype and carry out comprehensive experiments to evaluate its performance. For six basic emotions, the results show that EmoTracer can achieve 95.6% accuracy in intra-subject emotion classification and 80.7% accuracy with 5 shots in cross-subject emotion classification.