Many industrial and medical scenarios today are building on mobile devices as their end systems. These devices are typically connected to backend systems via the mobile phone network with their corresponding data services. Today's network protocols are optimized for reliable transfer of voice traffic, however, available bandwidth and latency for data traffic may vary significantly with location and motion speed of a mobile device. Within this paper, we report on experiences from our eHealth project "Fontane", where real-time streaming data from electrocardiographic devices has to be transferred from patient's home to doctors at a telemedicine center. In order to deal with varying bandwidth and transmission characteristics on the radio link, we propose a predictive model that allows for pro-active application reconfiguration in order to adapt to anticipated bandwidth variations. Our model is being integrated into a self-adaptive middleware for mobile communication. We present our initial study on network bandwidth estimation, which reveals that the measurement of the available bandwidth using standard end-to-end methods does not work well for mobile cellular networks due to a fairly high number of interference factors in wireless environments. We therefore propose a multilevel forecast model to better predict the network bandwidth of the immediate future.