Wearable photoplethysmogram (PPG) sensors are extensively used for remote monitoring of blood oxygen level and flow rate in numerous pervasive healthcare applications with diverse quality of service requirements. These sensors operate under severe resource constraints and communicate over an adverse wireless channel with human body-induced path loss and mobility-caused fading. In this paper, we take a generative model-based data collection approach towards achieving energy-efficient and reliable PPG monitoring. We develop two models that can generate synthetic PPG signals given a set of input parameters. These generative models are then used to design and implement a resource-efficient, reliable data reporting method for wireless PPG sensors. We investigate the performance of our method under realistic wireless channel error models and provide methods to improve accuracy at a marginal energy cost. We implement the proposed technique using existing sensor platforms and evaluate its performance on two datasets: the MIMIC database and data collected using commercial wearable sensors. Results for wearable sensor-based data show bandwidth and communication energy savings of 300:1, while maintaining a diagnostic accuracy above 94%.