Sensor networks provide a concise picture of complex systems and have been widely applied in health care domain. One typical scenario is to deploy sensors at different locations of human body and analyze the sensor measurements collectively to perform diagnosis of diseases. In this work, we are interested in differentiating peripheral arterial disease (PAD) patients from healthy people by monitoring peripheral blood pressure waveforms using electric sensors. PAD is an important cause of heart disease, which causes no significant symptoms until in a late stage. Therefore its early detection is of significant clinical values. Currently, PAD diagnosis either require large equipment or complicated, invasive sensor deployment, which is highly undesired in terms of medical expenses and safety considerations. To solve this problem, we present a novel approach to address the issue of high deployment cost in PAD detection via sensor networks. Assuming we are given many possibilities for sensor placement, each with different deployment cost, our goal is to select a small number of sensors with minimal costs while delivering accurate diagnosis. We solve this problem by treating each sensor as a feature, and designing a budget-constrained feature selection scheme to choose a compact, optimal subset of sensors, inducing very low deployment cost in terms of invasive treatment, while giving competitive classification accuracy compared with state-of-the-art feature selection method.