2026-10-14T18:08:14Z https://www.scidb.cn/oai

10.57760/sciencedb.30661 2026-03-18T17:03:44Z

2026-03-18 FEEL: A fNIRS–EEG Emotion Dataset and Benchmark Library doi:10.57760/sciencedb.30661 en The FEEL dataset is a publicly available, multimodal neurophysiological resource designed for affective computing and emotion recognition research. It features synchronously recorded electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) signals from participants experiencing music-evoked emotions. Developed to address the scarcity of standardized EEG–fNIRS datasets in the field, FEEL provides high-quality neural recordings, rich multi-dimensional emotion annotations, and a rigorously controlled experimental protocol—offering a reliable foundation for systematic studies on multimodal fusion strategies, emotion recognition algorithms, and neurovascular coupling in affective processing.Participants:34 healthy, right-handed volunteers (17 female, 17 male), aged 19–26 years (mean = 20.2, SD = 1.8), with normal or corrected-to-normal vision and no history of neurological or psychiatric disorders.Stimuli:38 emotionally evocative music clips selected from the internationally validated Soundtrack database, representing five discrete emotion categories: Happy, Anger, Fear, Sad, and Tender. Clips averaged ~17 seconds in duration (range: 13–26 s).Recorded Signals:EEG: 32-channel recording using Ag–AgCl active electrodes placed according to the international 10–20 system, sampled at 1000 Hz.fNIRS: 18-channel continuous-wave system covering prefrontal and temporal regions, using dual wavelengths (760 nm and 850 nm), sampled at 11 Hz, with a source–detector distance of approximately 30 mm.Synchronization: Millisecond-level alignment of EEG and fNIRS streams achieved via hardware triggers.Annotations:After each music clip, participants provided subjective ratings across multiple dimensions:Discrete emotion label (Happy, Anger, Fear, Sad, Tender);Dimensional scores for Valence, Arousal, and Dominance on a 1–9 Self-Assessment Manikin (SAM) scale;Liking score on a 1–9 scale (higher values indicate greater preference).Experimental Protocol:The experiment was conducted in a quiet, controlled environment. Each trial began with a 5-second eyes-closed rest period, followed by a 5-second white-noise baseline, music playback, and a self-paced emotional rating phase. The 38 trials were divided into two blocks separated by a 2-minute break, with trial order randomized per participant to minimize sequence effects. The study was approved by the Ethics Committee of the Affiliated Brain Hospital of Guangzhou Medical University (Approval No. 2026-004), and all participants provided written informed consent.Signal Preprocessing:All data underwent rigorous preprocessing to ensure signal quality:EEG: Powerline notch filtering, 1–100 Hz bandpass filtering, artifact removal via Infomax independent component analysis (ICA), and average re-referencing.fNIRS: Optical intensity converted to oxygenated (HbO) and deoxygenated (HbR) hemoglobin concentration changes using the modified Beer–Lambert law, followed by 0.01–0.2 Hz bandpass filtering, wavelet-based motion artifact correction, detrending, and z-score normalization. Affective computing; Emotion recognition; EEG; fNIRS; Multimodal dataset; Music-evoked emotion Wenhao Li Kai Li Ziqian Bai Yixuan Zeng Haolun Li Xuecheng Wu Guanchu Chen Ji Yuan Xie Shibo Zhang Jing Zhou Fengchun Wu Kai Wu RESTRICTED dataset Science Data Bank