Human activity recognition (HAR) is vital in healthcare settings, facilitating personalized care and early intervention. Recent advancements in sensor technology and machine learning have paved the way for more accurate and reliable HAR systems. However, the integration of multimodal data and the extraction of informative features remain significant challenges. In this paper, we propose a novel data fusion approach that leverages deep, nonlinear interactions between static and dynamic modalities to enhance the performance of HAR systems. Our method combines three types of interaction features derived from static data, latent representations learned by deep neural networks, and statistical features extracted from time series data. We explore two data fusion strategies: feature-level fusion and decision-level fusion. We conduct extensive experiments on six publicly available datasets, demonstrating the superiority of our approach compared to traditional methods. Our research highlights the importance of capturing complex interactions between multimodal data and opens up new possibilities for the development of advanced HAR systems in healthcare applications.