Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Discriminating coughs in a multi-bed ward environment
Thesis (MEng)--Stellenbosch University, 2021.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | en_ZA |
| Published: |
Stellenbosch : Stellenbosch University
2021
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867614104732041216 |
|---|---|
| access_status_str | Open Access |
| author | Leng, Corwynne |
| author2 | Niesler, Thomas |
| author_browse | Leng, Corwynne Niesler, Thomas |
| author_facet | Niesler, Thomas Leng, Corwynne |
| author_sort | Leng, Corwynne |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (MEng)--Stellenbosch University, 2021. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/123736 |
| institution | Stellenbosch University (South Africa) |
| language | en_ZA |
| last_indexed | 2026-06-10T12:46:44.579Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2021 |
| publishDateRange | 2021 |
| publishDateSort | 2021 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/123736 Discriminating coughs in a multi-bed ward environment Leng, Corwynne Niesler, Thomas Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering. Multi-Bed Ward Environment UCTD Cough -- Detection Neural networks (Computer science) Automatic speech recognition Accelerometers UCTD Thesis (MEng)--Stellenbosch University, 2021. ENGLISH ABSTRACT: Automatic cough detection algorithms play a key role in cough monitoring systems. These systems assess a patient’s recovery by monitoring the frequency and other characteristics of their coughs during treatment. An audio-based cough detection system evaluates features extracted from an audio signal and classifies these as either a cough or non-cough. However, these systems can struggle to distinguish between coughs of different individuals, such as a ward with more than one patient. This study designs and tests a cough detection algorithm that can reliably detect cough sounds in such an environment. We hypothesise that including the vibrations of the patient’s bed with the audio features will allow the classifier to differentiate between coughs originating from different beds in a hospital ward. Two datasets were compiled, containing simultaneously captured audio and accelerometer signals recorded by devices attached to the frame of each bed in a ward. These datasets were manually annotated with labels representing a cough from the bed in question, a cough from another source, and noise. The first dataset was used to train the classifier while the second was used to measure its performance. We extracted audio and accelerometer feature vectors using methods that have proved effective in automatic speech recognition systems, such as mel filter bank energies and mel-frequency cepstral coefficients. For the accelerometer signals, we included time-domain features. These feature vectors were presented to several classifiers, including convolutional neural networks and deep neural networks. Classifier training employed nested cross-validation to compensate for the small size of the training dataset and to allow for robust hyperparameter optimisation. The best classifier using audio and accelerometer features achieved an area under the receiver operating characteristic curve (AUC) score of 0.9842 while the best classifier using only audio features achieved an AUC score of 0.9334. Therefore, the inclusion of accelerometer features increased the AUC score by 0.0508 and improves the classifier’s ability to reject coughs from other sources by 10.72%. Additionally, the accelerometer features reduce the false detection of coughs from other sources and noise from 21.08% to 6.83%, while maintaining a sensitivity of 95%. We conclude that including the accelerometer signals of the patient’s bed with the audio features allows the classifier to better reject coughing sounds from sources other than the patient being monitored. AFRIKAANSE OPSOMMING: Outomatiese hoesbepalingsalgoritmes speel ’n sleutelrol in hoesmoniteringsisteme. Hierdie sisteme bepaal ’n pasiënt se herstel deur die frekwensie en ander eienskappe van hul hoesgeluide tydens behandeling te monitor. ’n Klankgebaseerde hoesbepalingsisteem evalueer kenmerke wat geëkstraeer is uit ’n klanksein en klassifiseer dit as ́of ’n hoes ́of ’n nie-hoes. Hierdie sisteme kan egter sukkel om tussen hoesgeluide van verskillende individue soos in ’n hospitaalsaal, waarin daar meer as een pasiënt is, te onderskei. Hierdie studie behels die ontwerp en evaluering van ’n hoesbepalingsalgoritme wat hoesklanke in so ’n omgewing op ’n betroubare manier kan bepaal. Ons veronderstel dat deur die vibrasies van ’n pasiënt se bed by die klankkenmerke van ’n hoes in te sluit sal die klassifiseerder in staat stel om te onderskei tussen hoesgeluide wat van verskillende beddens in ’n saal afkomstig is. Twee datastelle was saamgestel van klank - en versnellingsmeterseine wat gelyktydig opgevang is deur toestelle wat aan beddens in ’n hospitaalsaal gekoppel was. Hierdie datastelle was handmatig met etikette geannoteer wat aandui dat ’n hoes van ́of ’n bepaalde bed afkomstig was, ’n ander oorsprong het, ́of agtergrondsgeraas voorstel. Die eerste datastel was gebruik om die klassifiseerder op te lei terwyl die tweede een gebruik was om sy prestasie te meet. Klank- en versnellingsmeterkenmerkvektore was geëkstraeer deur van metodes gebruik te maak wat doeltreffend toegepas was in spraakherkenningsisteme soos die mel filterbank energieë en mel-frekwensie kepstrale ko ̈effisiënte. Tydsgebiedskenmerke was ingesluit by die versnellingsmeterseine. Hierdie kenmerkvektore was aan verskeie klassifiseerders, insluitend konvolusionêre en diep neurale netwerke, voorgestel. Klassifiseerderopleiding het gebruik gemaak van geneste kruisvalidasie om te vergoed vir die klein opleidingsdatastel en om ’n sterk of robuuste hiperparameter optimisasie toe te laat. Die beste klassifiseerder wat klank- en versnellingsmeterkenmerke gebruik het, het ’n area onder die ontvanger bedryfskenmerkende kurwe (AUC) telling van 0.9842 behaal, terwyl die beste klassifiseerder, wat net klankseine gebruik het, ’n AUC telling van 0.9334 behaal het. Die insluiting van versnellingsmeterkenmerke het dus die AUC telling met 0.0508 verhoog en verbeter die klassifiseerder se vermoë om hoes van ’n ander oorsprong te verwerp met 10.72%. Verder verminder die versnellingsmeterkenmerke die valse opsporing van hoes van ’n ander bron en agtergrondsgeraas vanaf 21.08% tot 6.83%, terwyl sensitiwiteit teen 95% gehandhaaf word. Ons gevolgtrekking is dat deur die versnellingsmeterseine, afkomstig van die pasiënt se bed, by die klankseine te voeg die klassifiseerder beter in staat sal wees om hoesgeluide wat nie van die pasiënt wat gemonitor word afkomstig is nie, te verwerp. Masters 2021-10-20T05:51:35Z 2021-12-22T14:18:34Z 2021-10-20T05:51:35Z 2021-12-22T14:18:34Z 2021-12 Thesis http://hdl.handle.net/10019.1/123736 en_ZA Stellenbosch University 116 pages application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Multi-Bed Ward Environment UCTD Cough -- Detection Neural networks (Computer science) Automatic speech recognition Accelerometers UCTD Leng, Corwynne Discriminating coughs in a multi-bed ward environment |
| title | Discriminating coughs in a multi-bed ward environment |
| title_full | Discriminating coughs in a multi-bed ward environment |
| title_fullStr | Discriminating coughs in a multi-bed ward environment |
| title_full_unstemmed | Discriminating coughs in a multi-bed ward environment |
| title_short | Discriminating coughs in a multi-bed ward environment |
| title_sort | discriminating coughs in a multi bed ward environment |
| topic | Multi-Bed Ward Environment UCTD Cough -- Detection Neural networks (Computer science) Automatic speech recognition Accelerometers UCTD |
| url | http://hdl.handle.net/10019.1/123736 |
| work_keys_str_mv | AT lengcorwynne discriminatingcoughsinamultibedwardenvironment |