Artificial intelligence increasingly influences daily decision-making, yet the safety of neural networks – particularly Spiking Neural Networks (SNNs) – remains a critical challenge. This dissertation enhances SNN safety in time-series classification through selective prediction, where high-uncertainty inputs are rejected to reduce errors. Two novel methods are introduced: a loss-based monitor that identifies risky inputs without relying on output probabilities, and a time-aware Conformal Prediction (CP) approach that corrects temporal uncertainties overlooked by standard CP. These methods are benchmarked on four public datasets and Infineon’s radar-based gesture data against state-of-the-art baselines, namely Evidential Deep Learning (EDL) and Neural Network Ensembles (NNEs). In addition to fixed-length evaluations, the research addresses real-time scenarios with streaming data, deploying SNNs via sliding windows and stateful architectures, where rejections act as safety interventions. Custom metrics are developed to better evaluate performance in this dynamic setting. Results reveal stateful deployment enables to catch more errors than sliding windows. EDL excels in accuracy-safety trade-offs, while CP delivers safety improvements with minimal computational overhead, offering scalable solutions for real-world SNN applications.