Detecting Patent Ductus Arteriosus in Neonatal Phonocardiograms

Mohammad Ali Zamani1, Ethan Grooby2, James Lacey3, Atul Malhotra3, Faezeh Marzbanrad3
1University of Tehran, 2University of British Columbia, 3Monash University


Abstract

Context: During fetal development, ductus arteriosus allows blood ejected by the right ventricle to bypass the lung. The persistence of this ductus after birth however is considered pathological and is termed patent ductus arteriosus (PDA).

Aims: Using phonocardiograms acquired from a digital stethoscope, the objective is to develop a classifier to detect PDA, as a means of an affordable screening tool.

Methods: Phonocardiogram recordings were obtained from 40 neonates, yielding 48 recordings. 17 recordings (15 neonates) had PDA, 8 recordings (7 neonates) had other heart structure abnormalities identified during echocardiography and were excluded in this study, and 23 recordings (21 neonates) had no heart structure abnormalities based on clinical assessment. Recordings were denoised using non-negative matrix co-factorisation to remove the lung, respiratory support and other noises. De-noised phonocardiograms were then segmented into S1, systolic, S2, and diastolic periods. Temporal (durations and maximum values), statistical (variance, skewness, and kurtosis), and power (total power, and 25-100Hz, 25-45Hz, 45-80Hz, 100-200Hz, and 200-400Hz band powers) were obtained for each heart segment and relative ratios between segments calculated. Feature selection was performed using a minimum redundancy maximum relevance algorithm, with the top four features used for training to prevent overfitting. The cost function was chosen for usage as a screening tool, with PDA misclassification cost twice the healthy phonocardiogram misclassification cost. Then, a support vector machine classifier was trained.

Results: Leave-one-subject-out cross-validation results showed a sensitivity of 88.2% and a specificity of 52.2%. The two misclassified PDA cases were assessed as small and moderate based on echocardiography assessment. Regular top features included S1, systolic, diastolic, and S1/S2 durations, max systolic value and max diastolic/S2 value.

Conclusion: Preliminary results show the promise of digital stethoscopes as a screening tool for PDA, prior to echocardiography, however, further research with a larger cohort is needed to confirm validity.