Tinybionet: A lightweight time-frequency network for biomedical signal classification on edge devices

Real-time analysis of biomedical time-series on edge devices faces significant challenges due to signal non-stationarity, environmental noise, and the strict computational limitations of embedded hardware. While deep learning offers superior accuracy, standard architectures are often too computationally demanding for embedded deployment. In this work, we propose TinyBioNet, a highly compact convolutional neural network designed explicitly for the efficient affective state classification on devices with limited resources. Unlike traditional 1D CNNs that implicitly learn spectral features, or lightweight architectures that treat the STFT as an isolated preprocessing operation, our approach natively embeds a fixed-basis time-frequency transformation directly within the network execution graph using parallel 1D convolutional layers. This design eliminates separate preprocessing stages, reduces runtime, and allows a highly compact backend 2D CNN topology to operate immediately on structured complex time-frequency feature maps. By combining this structured input with depthwise convolutions and residual connections, TinyBioNet achieves state-of-the-art performance with only 5.6k parameters. We further optimize the model for embedded targets through aggressive low-bit quantization. Comprehensive evaluations on three public datasets demonstrate robust generalization across various biomedical signal modalities. The proposed framework achieves classification accuracies of up to 98.89% and 99.38% for PPG and ACC signals, respectively, while maintaining negligible performance degradation under 4-bit integer quantization.