In the era of Internet of Things (IoT), an increasing amount of sensors is being integrated into intelligent wearable devices. These sensors have the potential to produce a large quantity of physiological data streams to be analyzed in order to produce meaningful and actionable information. An important part of this processing is usually located in the device itself and takes the form of embedded algorithms which are executed into the onboard microcontroller (MCU). As data processing algorithms have become more complex due to, in part, the disruption of machine learning, they are taking an increasing part of MCU time becoming one of the main driving factors in the energy budget of the overall embedded system. We propose to integrate such algorithms into dedicated low-power circuits making the power consumption of the processing part negligible to the overall system. We provide the results of several implementations of a pre-trained physical activity classifier used in smartwatches and wristbands. The algorithm combines signal processing for feature extraction and machine learning in the form of decision trees for physical activity classification. We show how an in-silicon implementation decreases up to 0.1 µW the power consumption compared to 73 µW on a general-purpose ARM's Cortex-M0 MCU.