Compressive sensing (CS) is a promising technology for realizing energy-efficient wireless sensors for long-term health monitoring. In this paper, we propose a data-driven CS framework that learns signal characteristics and individual variability from patients' data to significantly enhance CS performance and noise resilience. This is accomplished by a co-training approach that optimizes both the sensing matrix and dictionary towards improved restricted isometry property (RIP) and signal sparsity, respectively. Experimental results upon ECG signals show that our framework is able to achieve better reconstruction quality with up to 80% higher compression ratio (CP) than conventional frameworks based on random sensing matrices and overcomplete bases. In addition, our framework shows great noise resilience capability, which tolerates up to 40dB higher noise energy at a CP of 9 times.