Wearable-Based Emotion Recognition Using Electrocardiogram and Galvanic Skin Response and Instrumentation Audit: A Systematic Review

Simple and unobtrusive wearables, particularly watches and wristbands, now provide continuous monitoring with no inconvenience and are widely used by elderly people. Effective emotion recognition relies not only on advances in models but also on methods for signal acquisition and specification. In response to this imperative, we performed a systematic review of emotion recognition utilising wearable physiological signals, focusing on two practical modalities: electrocardiography (ECG), which indicates cardiac dynamics and emotional arousal, and galvanic skin response (GSR), a measure of electrodermal activity (EDA) reflecting sympathetic nervous system activation and conveniently recordable at the wrist, which reflects sympathetic nervous system activation and can be conveniently recorded at the wrist. We examined four prominent publishers (ScienceDirect, SpringerLink, Taylor & Francis Online, IEEE Xplore) and identified 549 papers; after excluding 90 review articles and 75 papers focused solely on electroencephalography (EEG), 384 were subjected to thorough screening. After excluding 20 studies focused solely on stress and 316 papers without ECG/GSR data, 48 studies utilising machine learning were identified (2019–2025). Traditional models like support vector machines remain prevalent, though deep learning techniques, particularly convolutional neural networks (CNNs) and hybrids, are more effective at discerning complex temporal patterns. Simultaneously, we conduct an instrumentation and measurement (I&M) audit on commonly utilised datasets. Protocols and sampling rates are generally documented; however, essential measurement details are frequently absent: electrode type and positioning, skin–electrode impedance, signal quality indices (SQIs), calibration, repeatability, and interdevice comparability. These limitations hinder equitable performance attribution (sensing versus modelling), limit uncertainty assessment, and complicate hardware transfer and cross-study evaluation. Standardising I&M metadata in public releases, when combined with AI evaluation practices, improves comparability and dependability in healthcare predictions based on wearable physiological information.