Despite the low-sampling rate of signals in medical sensor networks, a large number of measurement channels and severely energy-constrained devices could pose bandwidth and communication energy limitations. Compressive sensing is an efficient method of representing the data, which may help alleviate some of these data constraints. Although compressive sensing enables low-energy data reduction on the sensors, reconstruction costs are severe. The high reconstruction complexity typically pushes signal analysis to a base station. For many emerging medical applications, however, we want to aggregate and analyze the data on some local processing node. In this talk, I will describe a methodology to transform linear signal processing computations to the compressed domain so that the transformed computations can be applied directly to compressed data. I will present two example case studies where compressively-sensed spikes and electroencephalograms (EEGs) are directly analyzed to estimate neuron firing rates and to detect epileptic seizures, respectively. Besides circumventing reconstruction costs, this approach also provides a new power management knob by reducing computational energy with the number of input samples. I will show how we can take advantage of this power management knob in a custom IC fabricated in 130nm low-power CMOS. Host: Lin Zhong