Towards an unsupervised device for the diagnosis of childhood pneumonia in low resource settings: Automatic segmentation of respiratory sounds

Abstract

Pneumonia remains the worldwide leading cause of children mortality under the age of five, with every year 1.4 million deaths. Unfortunately, in low resource settings, very limited diagnostic support aids are provided to point-of-care practitioners. Current UNICEF/WHO case management algorithm relies on the use of a chronometer to manually count breath rates on pediatric patients: there is thus a major need for more sophisticated tools to diagnose pneumonia that increase sensitivity and specificity of breath-rate-based algorithms. These tools should be low cost, and adapted to practitioners with limited training. In this work, a novel concept of unsupervised tool for the diagnosis of childhood pneumonia is presented. The concept relies on the automated analysis of respiratory sounds as recorded by a point-of-care electronic stethoscope. By identifying the presence of auscultation sounds at different chest locations, this diagnostic tool is intended to estimate a pneumonia likelihood score. After presenting the overall architecture of an algorithm to estimate pneumonia scores, the importance of a robust unsupervised method to identify inspiratory and expiratory phases of a respiratory cycle is highlighted. Based on data from an on-going study involving pediatric pneumonia patients, a first algorithm to segment respiratory sounds is suggested. The unsupervised algorithm relies on a Mel-frequency filter bank, a two-step Gaussian Mixture Model (GMM) description of data, and a final Hidden Markov Model (HMM) interpretation of inspiratory-expiratory sequences. Finally, illustrative results on first recruited patients are provided. The presented algorithm opens the doors to a new family of unsupervised respiratory sound analyzers that could improve future versions of case management algorithms for the diagnosis of pneumonia in low-resources settings.

Ultrasound image intelligent diagnosis in community-acquired pneumonia of children using convolutional neural network-based transfer learning

“…The algorithm had a sensitivity of 94% and a specificity of 75% for the diagnosis of pneumonia. Other studies segmented the expiratory and inspiratory phases of breath sounds for parallel acoustic analysis ( 31 ). Sometimes patients don't have typical clinical manifestations.…”

Artificial intelligence-driven wearable technologies for neonatal cardiorespiratory monitoring. Part 2: artificial intelligence

“…35) employed the SVM algorithm for wheeze detection on the audio data recorded using smartphones in the paediatric population. In 2016, Sola et al (36) proposed to use traditional machine learning-based algorithms (Gaussian mixture model (GMM) and HMM) on the Mel-frequency filter bank from audio signals obtained from the digital stethoscope to detect childhood pneumonia. The GMM helps learn the unsupervised pattern of data, whereas the HMM helps find the sequential pattern of data.…”

Digital stethoscopes in paediatric medicine

“…In one study, telemedicine consultations for the monitoring of children with asthma via DS achieved comparable levels of satisfaction and outcomes in terms of asthma control when compared with in‐person consultations . In the near future, the DS may become a device able to aid the diagnosis of pneumonia in children in low‐resource settings, through the development of advanced computer algorithms (which is already underway) .…”

“…One such ANN was able to obtain 100% sensitivity and specificity for correctly distinguishing between innocent and pathological murmurs , while an ANN designed to detect adventitious respiratory sounds obtained an accuracy of 95.12% . ANNs to detect specific diagnoses such as pneumonia are currently being trialled . The definitive development of robust algorithms would strengthen the benefits of ANN and DS as an interpretation tool in routine clinical practice.…”