Spiking Neural Networks for Cross-Modal Medical Diagnostics: A Comprehensive Survey of Architectures, Learning Mechanisms, and Applications

Chronic epilepsy patients face a 2-3 times higher risk of stroke compared to the general population, yet continuous monitoring for early cerebrovascular events remains challenging due to the high energy demands of traditional artificial intelligence systems and strict privacy requirements of medical data. Spiking Neural Networks (SNNs) have emerged as the third generation of neural network models, offering significant advantages in energy efficiency, biological plausibility, and spatiotemporal information processing compared to traditional artificial neural networks. SSNs offer a promising alternative through asynchronous, spike-based computation that achieves significant energy efficiency. This survey systematically reviews the fundamentals of SNN architectures, learning mechanisms including Spike-Timing-Dependent Plasticity and surrogate gradient methods, and their applications across biomedical domains. This paper also focuses on fundamental neuron models including Hodgkin-Huxley, Leaky Integrate-and-Fire, and Izhikevich models, along with encoding strategies such as rate coding and temporal coding. Significant findings reveal that SNNs achieve comparable accuracy to traditional deep learning models while consuming 10-100× less energy, making them particularly suitable for wearable healthcare applications and edge computing. However, critical gaps exist in cross-modal spiking attention mechanisms, sparsity-aware federated learning protocols, and temporal explainability methods for clinical validation. Applications include EEG-based emotion recognition, ECG arrhythmia classification, EMG gesture recognition, and MRI-based disease detection, with emerging opportunities in integrated stroke detection for epilepsy patients.