Quality Assessment of Maternal and Fetal Cardiovascular Sounds Recorded from the Skin near the Uterine Arteries During Pregnancy

Dagbjört Helga Eiríksdóttir1, Rasmus Gundorff Sæderup2, Diana Riknagel2, Henrik Zimmermann2, John Hansen3, Johannes Struijk3, Samuel Emil Schmidt3
1Department of Health Science and Technology, Aalborg University, 2Viewcare A/S, 3Aalborg University


Introduction: Identifying and monitoring cardiovascular abnormalities during pregnancy is of high importance for the health and growth of the fetus. Cardiovascular auscultation of the pregnant abdomen is a novel non-invasive method for obtaining information on the maternal and fetal health including blood flow to the placenta, in both audible and infrasonic frequencies. However, the quality of such signals is often contaminated by ambient or unrelated abdominal sounds. It is therefore imperative to automatically identify and select the high-quality segments during the recording session for future analysis. This study aims to define and automatically distinguish high-quality signals.
Methods: A database of 324 recordings obtained from two microphones placed bilaterally on the abdomen of 90 pregnant women (gestational age of 28-41 weeks), with duration varying from 30s-180s, was used. The signals were bandpass filtered to infrasonic frequencies (2.5Hz-25Hz), divided into 10s segments (in total 3076) and areas with noise spikes were removed. The signal quality was assessed by visual inspection of the filtered segments and their autocorrelations. A good-quality segment was determined from the presence of detectable rhythm corresponding to the heart beat of either mother or fetus, and a good autocorrelation with repetitive and high peaks. Six features were calculated for segments with minimum five continuous seconds remaining after spike removal. A logistic regression model was trained and tested using the six features and used for classification. Results: Manual quality assessment resulted in 1007 good-quality segments. After spike removal, quality indices were calculated from 2457 segments. With an accuracy of 92.6% the classifier correctly categorized 906 good-quality segments out of 1007, corresponding 90.0% sensitivity and 1368 poor-quality segments out of 1450, corresponding to 94.3% specificity.
Conclusion: The results demonstrate the feasibility of automatically distinguishing good-quality segments from poor-quality segments at infrasonic frequencies.