MSCG-Net: A multi-scale cross-guidance neural network with generalizability for physiological signal analysis

• Proposing a multi-scale cross-guidance neural network with generality. • Integrating multiple receptive fields to aggregate detail and semantic information. • Combining distributed thinking to balance accuracy and efficiency. • Introducing sparse synaptic connectivity to optimize information interaction. • Achieving integrated deployment based on Android phones. Artificial intelligence (AI) plays a significant role in physiological signal analysis. Classification and regression are the two fundamental tasks of physiological signal analysis, and existing research primarily focuses on designing artificial narrow intelligence (ANI) for specific tasks. However, ANI requires continuous optimization of hyperparameters to adapt to different tasks and tends to have poor robustness. Meanwhile, the flexibility of integrating multiple ANIs to accomplish different tasks in actual deployment is limited. Therefore, this study proposes a general multi-scale cross-guidance neural network (MSCG-Net). It does not require re-optimization of hyper-parameters when applied to classification and regression, and it demonstrates favorable robustness when generalized to different tasks. Firstly, multi-scale mechanism is utilized to analyze physiological signals to extract features required for different tasks. Then, distributed architecture is used to optimize convolutional operators and network backbone to reduce the computational cost of the multi-scale mechanism. Finally, sparse synaptic connectivity is employed to enhance the information interaction among submodules in the distributed architecture to improve robustness. In various clinical scenarios, the classification accuracy and F1 score of MSCG-Net reach 99.41% and 99.28%, respectively, while the regression performance meets the British Hypertension Society A-grade standard and the Association for the Advancement of Medical Instrumentation standard. Furthermore, this study deploys MSCG-Net on an Android phone, with processing time of 12.88 s for analyzing 24 h of physiological signals. This research contributes to enhancing the generalizability of AI and optimizing its integrated deployment.