MULTI-STAGE ADAPTIVE ALGORITHM FOR HEART SOUND DETECTION AND RHYTHM ESTIMATION USING SERIAL ACOUSTIC TRANSDUCERS

Abstract

A multi-stage adaptive digital signal processing algorithm for phonocardiograms is proposed, targeted at serial acoustic transducers with a low signal-to-noise ratio, enabling affordable components to replace costly specialised medical sensors. It is shown that improving the class of the measurement chain has an analytical gain limit, because the in-band physiological interference cannot be removed by hardware; the refinement is therefore shifted from the hardware level to the algorithmic level. The algorithm resolves three problems left open in the authors’ previous hardware-oriented works: octave ambiguity of autocorrelation period estimation caused by two acoustic events per cardiac cycle; rhythm non-stationarity within a record; and identification of S1 within an unresolved S1/S2 pair. The ambiguity index  is introduced, and the multiple-lag threshold 0.45 is shown to correspond to a tone amplitude ratio of 0.24…4.21, which covers the physiologically possible range. For the windowed autocorrelation, a two-sided bound on the window length is derived, both bounds intersecting near 10 s under typical conditions, which justifies the window choice without empirical tuning. Tone detection is performed on an ensemble envelope combining the Shannon and Hilbert transforms, with a locally adaptive threshold based on the median and the interquartile range. The common rule «keep the louder peak» is shown to inject a jitter of 111…157 ms into the inter-beat interval series at 5…10 % mislabelling rates, an order of magnitude above physiological variability; a temporal-priority rule with a 1.8 dominance barrier is used instead. Rhythmogram cleaning is complemented by a safeguard against the false «normalisation» of arrhythmia. Decimation of the analytic path to 1 kHz reduces the computational cost 44-fold while adding an error of only 0.41 ms — 0.04 % of the variance of a typical tachogram. The total complexity of the algorithm is  at under 3 % load of a Raspberry Pi 3 core, making it suitable for edge computing and real-time telemedical monitoring.