163 - Novel, noninvasive vibroacoustic signature patterns of infant lungs during early development

Poster Number: 163
Publication Number: 163.2884

Background: Throughout the history of medicine and science, new measurement techniques have inexorably led to new breakthroughs. There has been no recent development of measurement, diagnostic or prevention tools of common conditions in preterm infants. Although sounds and palpitation of lungs, heart, and intestine were used as fundamental indicators of disease in the past, modern medicine is relying more and more on imaging, molecular testing, and blood-based biomarkers. We have embarked on a novel approach of capturing, cataloging, and interpreting the wide spectrum of inaudible and audible sounds and vibrations produced by different organs to screen, detect, diagnose, and monitor health and disease.

Objective: We used machine learning algorithms and a non-invasive instrument (iMPulse TORTM-Fig 1) that records sounds of infra- to ultrasonic frequencies to discern lung sounds of preterm and infants.

Design/Methods: In this IRB-approved study, 13 term and 8 preterm infants have been enrolled to date. Up to 24 longitudinal recordings (30 seconds each on right and left chest) have been performed in subjects staying in the NICU for an extended period. Assessments were compared to lung ultrasound scores performed at the same times. Data presented include those from first three (two full-term [NICU001 & 002] and one 26 wk gestation [003]) infants. Analysis of the remaining recordings and on new enrollees is continuing. For analysis, recordings were split into 5-second, non-overlapping segments, from which wavelet scattering coefficients were calculated and fed into a binary classification neural network to distinguish between full-term and pre-term [003] infants. Fig 2 represents t-distributed stochastic neighbor (t-SNE) embeddings of higher-dimensional feature vectors from out-of-sample test data that were extracted from the model.

Results: The out-of-sample test dataset had a classification accuracy of 73%, whereas the training dataset had 89% accuracy and the validation dataset had 87% accuracy. The 003 pre-term birth test data clustering in the t-SNE plot supports our hypothesis that the lung (and possibly heart) vibroacoustics are uniquely different from those in term infants.

Conclusion(s): These preliminary results provide a strong evidence that analysis of a larger dataset will yield even more robust results. When examined in conjunction with gestational age and known physiological changes during lung development, our model will provide unique vibroacoustic lung signatures for term and pre-term infants. Used longitudinally, such non-invasive assessments may predict BPD and other conditions during early NICU stay.