"A Lightweight Dual-Stream Framework for Stress Classification Using Multi-Scale ECG Fusion(Coming soon)"
Stress diagnosis typically relies on single- or continuous-cycle electrocardiogram (ECG) signals alone, which may fail to capture fine-grained beat-level morphology or long-term temporal dynamics. In this study, we propose DeepBeat-Rhythm FusionNet (DBRFNet), a lightweight dual-stream framework that integrates features from both single-cycle (beat) and continuous cycle (rhythm) ECG signals. By employing depthwise separable convolutions, an adaptive weight-sum fusion, and a Transformer Encoder, DBRFNet effectively captures morphological, relational, and temporal stress-related features with only 77.5 K parameters and 8.37 MFLOPs. Experimental evaluations on four public datasets (WESAD,MAUS,CLAS,andSWELL-KW)demonstratethatDBRFNet consistently outperforms single-stream approaches, achieving up to 99.40% accuracy in multiclass stress classification tasks. Real-world inference experiments on the Raspberry Pi 4B confirmed near-real-time performance with minimal latency of 1.17–2.27 ms. Furthermore, post-training quantization analysis showed that dynamic range quantization reduces model size by 57.6% (to 148 KB) while preserving diagnostic accuracy, indicating its high suitability for resource-constrained wearable edge devices. These results indicate that integrating single- and continuous-cycle ECG features improves diagnostic performance, and supports the use of DBRFNet as an efficient and lightweight framework for real-world stress monitoring.