Résumé
This paper presents a novel approach for automatic hemodynamic response detection in functional Near-Infrared Spectroscopy (fNIRS) signals, a modern neuroimaging technique that offers a portable and non-invasive solution for monitoring brain activity in naturalistic settings. We address the challenges posed by signal complexity and inter-subject variability by leveraging two entropy-based methods, Permutation Entropy (PE) and Phase Rectified Signal Average (PRSA), which focus on the statistical properties of noise rather than the signal content itself. Our experiments on raw, annotated fNIRS recordings data demonstrate that these methods achieve performance comparable to traditional machine learning algorithms, with the additional advantage of requiring no prior training. Their versatility, adaptability to various signal types, and significant reduction in computational time make them particularly well-suited for realtime applications in dynamic environments, further enhancing the practical potential of fNIRS in cognitive and clinical research.