| 日時(Date) |
Jul 1st, 2026 4th period (15:10-–16:30) |
|---|---|
| 場所(Location) | Room A607 / Online Lecture (Zoom link: https://zoom.us/j/9321514910?pwd=bzBDbGtCR2poWTNDb3ZXa0hEWkNGQT09) |
| 司会(Chair) | Kazushi Ikeda sensei |
| 講演者(Presenter) | Prof. S. Neelakandan, R.M.K. Engineering College, Chennai, India |
| 題目(Title) | Learned Time–Frequency Tokenizer with Temporal Transformer Using Relative Positional Encoding for Psychological Stress Detection from Phonocardiography Signals |
| 概要(Abstract) | Psychological stress induces subtle yet persistent alterations in cardiac dynamics that can be effectively captured through phonocardiography (PCG) signals. However, reliable stress detection from PCG remains challenging due to the non-stationary, multi-scale, and temporally varying nature of heart sounds. To address these challenges, this paper proposes an end-to-end attention-driven framework that integrates a Learned Time–Frequency Tokenizer (LTFT) with a Temporal Transformer employing Relative Positional Encoding (RPE) for accurate psychological stress detection. Unlike conventional approaches that rely on handcrafted time–frequency representations such as wavelet or constant-Q transforms, the proposed LTFT adaptively learns stress-discriminative spectral–temporal patterns directly from raw PCG signals using multi-scale convolutional filter banks. The resulting time–frequency tokens are subsequently processed by a Temporal Transformer, where relative positional encoding preserves temporal order while enabling effective modeling of long-range cardiac rhythm dependencies. This architecture jointly captures localized heart sound events and global stress-induced temporal variations. Experimental evaluations conducted on a publicly available PCG-based mental stress dataset demonstrate that the proposed model achieves high classification accuracy, outperforming conventional CNN-, TCN-, and LSTM-based baselines in terms of accuracy, precision, recall, and F1-score. The results confirm the effectiveness of adaptive representation learning combined with attention-based temporal modeling for robust and accurate PCG-based psychological stress assessment. |
| 講演言語(Language) | English |
| 講演者紹介(Introduction of Lecturer) | Professor S. Neelakandan is a Professor-Research in the Department of Computer Science and Engineering at R.M.K Engineering College, Chennai, India. He received the Brain Pool Research Fellowship during his Postdoctoral research. He has also served as a Visiting Professor at UCSI University and an Associate Editor in the research fields of AI, Data Science, Human Computer Interaction, and Robotics. His website is https://www.drsneelakandan.com/ |