Spiking Neural Networks (SNNs), and particularly Spiking Transformers, have recently emerged as a promising solution for energy-efficient processing of event-based sensor data in healthcare applications. However, existing architectures are inherently rigid, being trained and deployed as static networks with fixed parameter counts and computational graphs. This rigidity presents a critical bottleneck for deployment on neuromorphic hardware (e.g., Loihi, Spinnaker), where on-chip hardware limitations often force the use of smaller, lower-accuracy models that sacrifice representational capacity and accuracy. To bridge this gap, we introduce the Elastic Spiking Transformer, a novel architecture that embeds runtime adaptability directly into the spiking paradigm. Drawing inspiration from Matryoshka-style representation learning, our framework introduces nested elasticity within the Feature Extractor, Spiking Self-Attention, and Feed-Forward blocks. By training a single “universal” model with granularity-aware weight sharing, we enable dynamic slicing of network width and attention heads during inference without retraining. This elasticity offers a unique dual advantage in the SNN domain. First, it solves the memory-constrained problem by allowing the model to instantly scale its parameter footprint to fit diverse hardware specifications without retraining. Second, and crucially, we show that elasticity in SNNs not only lowers synaptic weight counts but also directly lowers the total spike firing rate. By dynamically reducing the number of active neurons, our approach achieves a proportional reduction in synaptic operations, unavailable to standard artificial neural networks. We validate our approach on standard benchmarks (CIFAR10/100, CIFAR10-DVS) and the novel EHWGesture clinical gesture understanding dataset. Our results demonstrate that a single Elastic Spiking Transformer spans a wide range of complexity trade-offs, matching or surpassing the accuracy of independently trained baselines while enabling adaptive, realtime gesture recognition on resource-constrained edge devices.