Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Deep learning based methods for tuberculosis cough classification
Thesis (MEng)--Stellenbosch University, 2023.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | en_ZA en_ZA |
| Published: |
Stellenbosch : Stellenbosch University
2023
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613888833388544 |
|---|---|
| access_status_str | Open Access |
| author | Frost, Geoffrey |
| author2 | Niesler, Thomas |
| author_browse | Frost, Geoffrey Niesler, Thomas |
| author_facet | Niesler, Thomas Frost, Geoffrey |
| author_sort | Frost, Geoffrey |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (MEng)--Stellenbosch University, 2023. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/127066 |
| institution | Stellenbosch University (South Africa) |
| language | en_ZA en_ZA |
| last_indexed | 2026-06-10T12:43:18.087Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2023 |
| publishDateRange | 2023 |
| publishDateSort | 2023 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/127066 Deep learning based methods for tuberculosis cough classification Frost, Geoffrey Niesler, Thomas Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering. Deep learning (Machine learning) Tuberculosis -- Diagnosis Nosology Thesis (MEng)--Stellenbosch University, 2023. ENGLISH ABSTRACT: Automated systems for disease identification have the potential to streamline the patient diagnosis process and provide insight to physicians. Exploratory studies have developed such systems for tuberculosis (TB) which rely on the cough audio signal produced by patients to determine, with the use of statistical classifiers, if they may have TB or not. Although these studies were small and the algorithms developed rudimentary, promising results were achieved. In this study, we build upon existing work and investigate the application of various deep learning -based approaches for TB cough classification. Since such systems would eventually be deployed in a multitude of environments and to increase the size of the dataset which is useful for model development, we combine the datasets used in previous studies. Multiple classifiers are developed, which include architectures based on bidirectional long short-term memory networks (BiLSTM), convolutional neural networks (CNN), attention and transformers. Additionally, the application of various large pre-trained models (ResNet, Audio Spectrogram Transformer and wav2vec2.0) to TB cough classification is investigated. Moreover, we develop a unique cough classification pretraining task to better initialise model parameters. Substantial classification performance improvements are observed compared to the previous best methods. In particular, the pre-trained BiLSTM architecture achieved relative improvements in AUC and EER of 9.33% and 64.67% respectively compared to the baseline system. More generally, the use of pre-training almost always improved the performance of these metrics, and always lead to better generalisation, observed by a reduction in metric standard deviation across evaluation sets. Due to the cross-validation procedure used during development, the choice of decision thresholds was sub-optimal, which subsequently lead to poor sensitivity and specificity. This was typically worse with pre-training. However, even when using oracle decision thresholds, classifiers were unable to reach the WHO standards for such a diagnostic tool. Additionally, we conduct a brief investigation into patient identity as a confounding factor during training and subsequent deep learning-based mechanisms to inhibit its learning. Whilst we present clear evidence that models learn the identity of patients in conjunction with the underlying TB signal, its removal does not significantly impact performance. Lastly, using insights from feature importance experiments, attention weights analysis, and adversarial synthesis we provide clues regarding the origin and characteristics of the learnt TB signal in cough. Specifically, we use these methods to identify the most important frequency bands for classification, the importance of certain temporal regions in cough, and the distinct spectral characteristics between idealised TB and non-TB coughs. AFRIKAANS OPSOMMING: Geautomatiseerde stelsels vir siekte-identifikasie het die potensiaal om die pasi¨entdiagnoseproses meer vaartbelyn te maak en insig aan dokters te verskaf. Verkennende studies het sulke stelsels vir tuberkulose (TB) ontwikkel wat staatmaak op die hoes-klank wat deur pasi¨ente geproduseer word om, met die gebruik van statistiese klassifiseerders, te bepaal of hulle moontlik TB het of nie. Alhoewel hierdie studies klein was en die algoritmes rudimentˆer ontwikkel het, is belowende resultate behaal. In hierdie studie bou ons op bestaande werk en ondersoek ons die toepassing van verskeie diepleer-gebaseerde benaderings vir TB-hoesklassifikasie. Aangesien sulke stelsels uiteindelik in verskeie omgewings ontplooi sou word en om die grootte van die datastel wat nodig is vir modelontwikkeling te vergroot, kombineer ons die datastelle wat in vorige studies gebruik is. Veelvuldige klassifiseerders word ontwikkel, wat argitekture insluit wat baseer is op “bidirectional long short-term memory networks” (BiLSTM), “convolutional neural networks” (CNN), “attention” en “transformers”. Daarbenewens word die toepassing van verskeie groot vooraf-opgeleide modelle (ResNet, Audio Spectrogram Transformer en wav2vec2.0) op TB-hoesklassifikasie ondersoek. Boonop dit, ontwikkel ons ‘n unieke hoesklassifikasie-voorafopleidingstaak om modelkomponente beter te inisialiseer. Aansienlike klassifikasie prestasieverbeterings word waargeneem in vergelyking met die vorige beste metodes. In die besonder het die vooraf-opgeleide BiLSTM-argitektuur verbetering behaal in AUC en EER van 9.33% en 64.67% onderskeidelik in vergelyking met die basislynstelsel. Oor die algemeen het die gebruik van vooropleiding byna altyd die prestasie van hierdie maatstawwe verbeter, en het altyd gelei tot beter veralgemening. Dit is waargeneem deur ‘n vermindering in metrieke standaardafwyking oor die evalueringsstelle. As gevolg van die kruisvalideringsprosedure wat tydens ontwikkeling gebruik is, was die keuse van besluitdrempels sub-optimaal, wat vervolgens gelei het tot swak sensitiwiteit en spesifiekheid. Dit was gewoonlik slegter met vooropleiding. Selfs wanneer orakelbesluitdrempels gebruik word, kon klassifiseerders egter nie die WGO-standaarde vir so ‘n diagnostiese hulpmiddel bereik nie. Daarbenewens doen ons ‘n kort ondersoek na pasi¨entidentiteit as ‘n verwarrende faktor tydens opleiding en daaropvolgende diepleer-gebaseerde meganismes om die leer daarvan te verhoed. Alhoewel ons duidelike bewyse aanbied dat modelle die identiteit van pasi¨ente leer in samewerking met die onderliggende TB-klanksein, het die verwydering daarvan nie ‘n noemenswaardige impak op prestasie nie. Laastens, deur gebruik te maak van insigte van kenmerkbelang-eksperimente, aandaggewigte-analise en teenstrydige sintese, verskaf ons leidrade aangaande die oorsprong en kenmerke van die aangeleerde TB-klanksein in ‘n hoes. Spesifiek, ons gebruik hierdie metodes om die belangrikste frekwensiebande vir klassifikasie te identifiseer, asook die belangrikheid van sekere temporale streke in ‘n hoes, en die duidelike spektrale kenmerke tussen ‘n ge¨ıdealiseerde TB en ‘n nie-TB hoes. 2023-03-03T09:52:22Z 2023-05-18T07:02:36Z 2023-03-03T09:52:22Z 2023-05-18T07:02:36Z 2023-03 Thesis http://hdl.handle.net/10019.1/127066 en_ZA en_ZA Stellenbosch University xx, 139 pages : illustrations. application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Deep learning (Machine learning) Tuberculosis -- Diagnosis Nosology Frost, Geoffrey Deep learning based methods for tuberculosis cough classification |
| title | Deep learning based methods for tuberculosis cough classification |
| title_full | Deep learning based methods for tuberculosis cough classification |
| title_fullStr | Deep learning based methods for tuberculosis cough classification |
| title_full_unstemmed | Deep learning based methods for tuberculosis cough classification |
| title_short | Deep learning based methods for tuberculosis cough classification |
| title_sort | deep learning based methods for tuberculosis cough classification |
| topic | Deep learning (Machine learning) Tuberculosis -- Diagnosis Nosology |
| url | http://hdl.handle.net/10019.1/127066 |
| work_keys_str_mv | AT frostgeoffrey deeplearningbasedmethodsfortuberculosiscoughclassification |