Classification of Selected Cardiac Abnormalities through Machine Learning.

Date
2022-04
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH SUMMARY: Cardiovascular diseases contribute to a large number of deaths worldwide per year. From an engineering perspective, an opportune point of intervention is the examination phase of a patient where the equipment and supporting software is concerned. This study aims to develop a prototype supervised machine learning algorithm that can be used as a diagnostic tool in medical practise. Four hundred and six (406) Echocardiography examinations were col- lected containing six (6) different cardiac abnormalities associated with the left ventricle and aortic valve. Data was considerably insufficient thus augmenta- tion techniques were required to generate synthetic samples. Image processing techniques and various calculations were used to derive measurements and fea- tures to be suitable input for the machine learning models. Random Forest and Neural Network models with a variety of dimensions were developed and trained in 3 different tests. The first 2 tests investigated the value of engi- neering (measurement-derived) and medical (patient information) features to model outputs. Test 3 investigated the effect of various training set si'.es. Both models were better informed by medical features than those extracted geometrically or calculated. This was found due to the effect of noise distort- ing measurements extracted for features. Models also performed better on the largest training set si'.e (90% of data). All models were evaluated by selected performance metrics and/or learning curves (where applicable). The most suitable model selected was a Random Forest instance, as Neural Networks were prone to overfitting training data. These results were not true refiections of either model's capabilities due to the underlying data representativeness issue.
AFRIKAANS OPSOMMING: Kardiovaskulere siektes dra by tot 'n groot aantal sterftes wereldwyd per jaar. Vanuit 'n ingenieursperspektief is die ondersoekfases van pasientbehandeling die venster vir intervensie. Die doel van die studie is om 'n prototipe toe- sig masjienleer-algoritme te ontwikkel wat as 'n diagnose hulpmiddel in die mediese praktyk gebruik kan word. Vierhonderd-en-ses (406) eggokardiogram toetse is ingesamel vir ses (6) verskillende kardiale abnormaliteite wat verband hou met die linker ventrikel en aortaklep. Uiters onvoledoende data is versa- mel; en daarom was dit noodsaaklik om van aanvullingstegnieke gebruik te maak om sintetiese monsters te genereer. Beeldverwerkingstegnieke en ver- skeie berekeninge is gebruik om afieidings kenmerke en metings te identifiseer om geskikte insette vir die masjienleermodelle te bied. Random Forest en Neu- rale Netwerk modelle met 'n verskeidenheid dimensies is ontwikkel en opgelei in 3 verskillende toetse. Die eerste twee toetse het die waarde van ingenieurs- wese (meting-afgeleide) en mediese (pasientinligting) kenmerke ondersoek om uitsette te modelleer. Die derde toets het die effek van verskeie opleidingstel groottes ondersoek. Beide modelle is beter ingelig deur mediese kenmerke as die wat meetkundig uitgewerk of bereken is. Die rede hiervoor is later ge"n- detifiseer as geraas metings wat as kenmerke onttrek is en gebruik is. Modelle het beter gevaar met die grootste opleidingstel grootte (90% van die data). Alle modelle is geevalueer deur prestasiemaatstawwe of leerkurwes (waar van toepassing). Die mees geskikte gekose model; was 'n Random Forest geval; aangesien Neurale netwerke geneig was om opleidingsdata te oorpas. Hierdie resultate was nie ware weerspieeling van enige van die modelle se vermoens nie as gevolg van die teenwoordigheid van onderliggende data kwessies en geraas.
Description
Thesis (MEng)--Stellenbosch University, 2022.
Keywords
Cardiovascular system -- Diseases, Echocardiography, Machine learning, UCTD
Citation