Hidden Markov Model-Based Asthmatic Wheeze Recognition Algorithm Leveraging the Parallel Ultra-Low-Power Processor (PULP)

Abstract

Asthmatic symptoms can be quantified by a wearable sensor system, recording respiratory sounds on patient’s skin surface, and performing automated asthmatic wheeze recognition based on time-frequency features. In order to enable long-term autonomy of such sensor system, a crucial design requirement is ensuring energy-efficient yet accurate wheeze recognition performance. We presented a Hidden Markov Model-based algorithm for recognition of wheezing intervals durations, by sequentially extracting individual wheezing-frequency lines from the spectrogram of respiratory sounds. In this paper we compare its implementation on an ARM Cortex-M4 processor and an emerging parallel ultra-low-power processing platform PULP Fulmine. It is shown that the algorithm enables wheeze recognition with 82.85% of sensitivity and 95.61% specificity, for only 0.9-1.6 mW of power. It is experimentally verified that algorithm benefits from a multi-core architectures such as PULP Fulmine. The implementation on this platform brings up to around 40% reduction of average power spent on processing, compared to the ARM Cortex-M4 Blue Gecko.