Research project: Time-structure based reconstruction of physiological sources extracted from noisy absominal phonograms

The abdominal phonogram is a signal recorded by a sensitive acoustic sensor positioned on the maternal womb. The signal conveys information which is valuable for antenatal foetal surveillance (e.g. heart sounds and foetal movements), but hidden by maternal and environmental noises. To recover such information, previous work successfully used Single-Channel Independent Component Analysis (SCICA) to decompose the phonogram into independent components (ICs). After that, knowing that some ICs belong to the same physiological process (i.e. maternal cardiovascular, maternal respiratory or foetal cardiac), ICs with similar spectral content were automatically grouped using K-means. This step, although achieved a fast and unsupervised classification, misclassified some foetal and maternal ICs. Consequently, the physiological sources recovered became distorted and virtually useless for studying foetal condition.

In this work, the rich time-structure of the physiological components underlying the abdominal phonogram is exploited to automatically classify similar components and thus, retrieve the independent sources corresponding to maternal activity (respiratory and cardiovascular), foetal heart sounds, and noise. To do so, a rhythmicity-based analysis was proposed. The scheme, based on autocorrelation and PSD analysis, was tested on a dataset composed of 750 ICs that were extracted from segments of 25 single-channel phonograms (recorded at foetal gestational ages between 29 and 40 weeks).

Results showed that, since this rhythmicity-based scheme is based on autocorrelation and PSD analysis, it manages not only to quickly and automatically group similar ICs, but also to correlate the recovered sources with specific physiological phenomena (either maternal or foetal), which is a desirable advantage. Further research will be conducted on more phonograms and explore alternatives for dimensional reduction and reconstruction of entire time-series suitable for surveillance, not only for foetal well-being but also for maternal condition.