Emotion detection unveiled: A cognitive-computational synthesis of physiological models, machine learning, and datasets.

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Descriptive accuracy improvement from 70–75% (traditional ML) to >95% (modern DL architectures including Transformers, self-supervised learning, diffusion models) — no pooled effect size calculated

• Primary Outcome: Emotion recognition accuracy using physiological signals (EEG, ECG, GSR) with modern deep learning architectures on benchmark datasets

• Nnt Or Sensitivity: Not applicable — no clinical NNT, sensitivity, or specificity data reported; accuracy figures are dataset-specific benchmark metrics without clinical diagnostic framing

• Confidence Interval: Not reported

Clinical Application

Current feasibility is limited to research and controlled settings. Real-time EEG, ECG, and GSR acquisition requires specialised equipment, signal processing expertise, and controlled environments. Consumer-grade wearable translation remains technically challenging. Cross-subject generalisability — a key limitation acknowledged by the review — must be resolved before clinical deployment is feasible. In the Australian context, physiological emotion recognition systems are not currently listed on the Australian Register of Therapeutic Goods (ARTG) as medical devices for clinical use. The TGA's Software as a Medical Device (SaMD) framework would apply to any clinical deployment. The RACGP does not currently provide guidelines for affective computing in primary care. Mental health applications may intersect with MBS telehealth item numbers and digital mental health platforms (e.g., Head to Health), but regulatory and clinical validation pathways remain undefined. The PBS does not fund any related technology. Australian researchers should note that benchmark datasets (SEED, DREAMER) were not developed with Australian population samples, raising questions about cross-cultural emotional expression validity. Potentially applicable to research contexts involving affective computing in mental health monitoring, neurological assessment, and human-computer interaction. Not yet validated for direct clinical use in any patient population.

Abstract

This comprehensive survey synthesizes state-of-the-art advancements in emotion recognition based on physiological signals, specifically focusing on the paradigm shift occurring between 2021 and 2025. Crucially, we move beyond a technical review by establishing a novel Cognitive-Computational Synthesis Framework (CCSF). This framework explicitly maps multimodal physiological manifestations (e.g., electroencephalogram (EEG), electrocardiogram (ECG), and galvanic skin response (GSR)) to underlying cognitive processes, such as attentional allocation, arousal regulation, and perceptual bias, providing a theoretical foundation for explainable AI (XAI) in affective computing. We meticulously examine the transition from traditional machine learning to advanced deep learning architectures, highlighting how recent innovations in Transformers, self-supervised learning, and diffusion models have shattered previous performance plateaus. While earlier dimensional models were often limited to 70-75% accuracy, this survey details how modern architectures now achieve benchmarks exceeding 95% on seminal datasets like SEED and DREAMER. Furthermore, the survey provides a rigorous analysis of 40 key studies (identified via PRISMA protocols), evaluating them based on their validation strategies, cross-subject generalizability, and adversarial robustness. By bridging the gap between raw physiological data and cognitive theory, this work offers a strategic roadmap for the next generation of robust, interpretable, and real-time emotion recognition systems.

References

1. 1.Machhi, V., & Shah, A. (2026). Emotion detection unveiled: A cognitive-computational synthesis of physiological models, machine learning, and datasets. Cognitive, Affective, & Behavioral Neuroscience. https://doi.org/10.1109/TAMD.2015.2431497 [Note: DOI metadata discrepancy identified — assigned DOI corresponds to a 2015 IEEE publication; independent verification recommended prior to citation]