An audio classification system for detecting pulmonary edema
Pulmonary edema is a condition where water engorges the alveolar beds. The water is pushed into air spaces and thus reduces normal oxygen movement. It can cause hemoptysis, difficulty in breathing, shorted of breath, and gurgling and wheezing sounds during breathing. It can be caused by congestive h...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
2019
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/90282 http://hdl.handle.net/10220/48518 |
| Summary: | Pulmonary edema is a condition where water engorges the alveolar beds. The water is pushed into air spaces and thus reduces normal oxygen movement. It can cause hemoptysis, difficulty in breathing, shorted of breath, and gurgling and wheezing sounds during breathing. It can be caused by congestive heart failure, kidney failure, high altitude exposure, lung damage, and major injuries.
Currently, physicians will carry out physical examinations before performing X-rays, CT-scans and electrocardiograms. Physicians mainly use auscultations to check for abnormal heart sounds, crackles, increased heart rate and rapid breathing. Most patients with pulmonary edema must face the disease and its implications for the rest of their lives. They must visit the physician for a check-up whenever they show symptoms, and if left untreated, they could suffer suffocation. These constant trips to the physician can add stress, worry and inconvenience. Physicians also in turn face a greater workload.
This thesis proposes new insights and methods to detect pulmonary edema using auscultation recordings. The first contribution makes use of feature engineering to gain new insights into the characteristics of these audio recordings. Empirical mode decomposition is used to improve the classification performance.
In the first contribution, it is shown that conventional feature selection algorithms are not designed to handle external interfering hospital noises. In the second contribution, improvements are made on the robustness with the introduction of a new system. The second contribution uses non-negative matrix factorization (NMF) to develop a robust audio classification system to detect pulmonary edema. A study was done to compare feature engineering approaches with feature selection techniques against NMF. Different NMF schemes were investigated and compared with Principal Component Analysis. Background noise collected from hospitals and speech from a speech corpus database were used to simulate noisy data. The system was then tested using noisy data. NMF was also used as a signal enhancement tool, which improved the classification scores.
Contributions in the third chapter are divided into two main parts. Inspired by the analysis of spectrograms of breath recordings, the first section investigates the use of Latent SSVM to show that breath sounds contain latent information that could potentially be exploited. To further improve the proposed NMF system, the second part of the contribution uses NMF to develop a cascading dictionary learning algorithm that allows the use of sample subsets to increase classification accuracy. The SVM classifier is used to group subsets of data and calculate dictionaries on each subset of data. These subsets can be thought of as latent variables that have been grouped into these subsets. The method is described and we compare results with benchmarks, including results from the previous chapter.
In conclusion, this thesis focuses on the automatic classification of pulmonary edema with the use of breath recordings through the chest wall. In particular, feature engineering was used to gain new insights. NMF was next used to develop a robust classification system. Lastly, Latent SSVM was used to show that breath sounds contain latent information that could be exploited. A cascading dictionary learning algorithm was proposed for that purpose and improved classification results in noisy conditions. |
|---|