Vision-based action recognition systems face significant challenges in adverse weather conditions, a limitation we encountered during preliminary experiments for elderly monitoring in Moroccan healthcare settings where fog and dust storms are common. This deployment barrier motivated our investigation into multimodal sensing approaches. This paper presents a multimodal network that combines micro-Doppler radar signatures with RGB-infrared vision for weather-robust action detection. Our approach introduces a weather-adaptive attention mechanism that dynamically balances radar and vision contributions based on environmental conditions. The proposed architecture features hierarchical cross-modal attention that processes micro-Doppler signatures using transformer-based temporal attention while simultaneously extracting weather-robust visual features through enhanced RGB-IR fusion. This cross-modal attention mechanism attempts to create unified weather-invariant representations that maintain coherence across environmental conditions where traditional vision systems experience significant degradation. Our evaluation suggests the system achieves 94.7% accuracy in clear weather while maintaining 87.5% in adverse conditions, a 7.2% degradation that, while non-trivial, seems acceptable for continuous healthcare monitoring where reliability arguably matters more than perfect accuracy. The system achieves real-time performance with 87 ms inference time on edge devices, enabling practical IoT healthcare deployment. This work addresses what we see as critical deployment barriers for elderly care monitoring while preserving privacy through radar’s non-identifiable sensing capabilities, though we recognize that privacy guarantees in multi-modal systems require ongoing scrutiny as fusion techniques become more sophisticated.