diplomsko delo
Nejc Mlakar (Author), Andrej Brodnik (Mentor), Janez Žibert (Co-mentor)

Abstract

Vsaka četrta smrt v Ameriki je posledica različnih vrst bolezni srca. Te bolezni v večini povzroči nepravilno prehranjevanje, pomanjkanje gibanja ali starost. Nekatere od teh bolezni pa so lahko tudi prirojene. Da zdravniki vidijo, ali je s srcem bolnika kaj narobe, morajo pregledati EKG signal, ki ga srce oddaja. EKG signali se razlikujejo po številu sledi, ki jih vsebujejo, sledi pa določajo, na keterih delih telesa se je meril EKG signal. Značilnosti EKG signala so QRS kompleksi ali tako imenovani vrhovi. Značilnice QRS kompleksov so oblika, število pojavitev ter razdalja od prejšnjega QRS kompleksa. Vse te značilnice nam lahko povedo, za kakšno vrsto bolezni srca gre. Največja težava je v tem, da so si QRS kompleksi zelo podobni med seboj, kar lahko privede do napačne diagnoze in zdravljenja bolnika. Cilj te diplomske naloge pa je, da bi s pomočjo strojnega učenja lažje, hitreje in bolj natančno določili za kakšno vrsto aritmije gre, kar bi lahko zdravnikom olajšalo delo. Za dosego cilja smo uporabili podatkovno zbirko, ki je vsebovala označen enodimenzionalni EKG signal 24 pacientov. Podatke smo razdelili na tri medsebojno izklučujoče množice (učno, validacijsko in testno), kot je v skladu s standardi. Globoka nevronska mreža je nato na testnih primerih poskušala napovedati, za kakšno vrsto bolezni gre. Ker je za nekatere bolezni premalo podatkov, je bil naš cilj kar se da pravilno klasificirati bolezni, za katere smo imeli dovolj podatkov. Globoka nevronska mreža je tako poskušala napovedati deset različnih vrst aritmij, pri čemer jih je pravilno klasificirala v 91,5 % primerov.

Keywords

globoko učenje;nevronske mreže;QRS kompleks;računalništvo;računalništvo in informatika;visokošolski strokovni študij;diplomske naloge;

Data

Language: Slovenian
Year of publishing:
Typology: 2.11 - Undergraduate Thesis
Organization: UL FRI - Faculty of Computer and Information Science
Publisher: [N. Mlakar]
UDC: 004.85:616.12-008.318(043.2)
COBISS: 1538354883 Link will open in a new window
Views: 691
Downloads: 175
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Classification of heart arythmias with deep neural networks
Secondary abstract: Every fourth death in America is a consequence of some sort of heart disease. These diseases are mostly caused by unhealthy eating habits, lack of moving and old age. Some people can be born with heart diseases already formed. Doctors have to analyze ECK signal that the heart emits in order to correctly determine the kind of heart disease. ECK signals differ from each other by the number of leads they have. Number of leads determine where on the patient's body ECK signal was recorded. Main feature of ECK signal are QRS complexes, sometimes in this document referred to as peaks. QRS complexes have a number of features. Most important are shape, frequency and length from previous peak. All of this features can tell us what kind of heart disease does the patient have. The biggest problem with analyzing QRS complexes is that most of them are very much alike, which could lead to the wrong diagnosis and treatment. The goal of this seminar was to use machine learning with deep neural network to better analyze QRS complexes, to the point where the machine could predict what kind of heart disease it is. This kind of AI could help doctors identify heart diseases faster and more accurately. To achieve this goal we used dataset that contained two-dimensional marked ECK singal from 24 patients. Dataset was split into three separate datasets (learning, validation and testing dataset) according to the practices in the industry. Neural network used a part of this dataset to predict what kind of heart arrhythmia does the signal contain. Since some of the diseases only had a few cases, we decided to only train and test our neural network on diseases that occurred more frequently. Our neural network tried classifying QRS complexes into 10 classes. It was successful in 91,5 % of cases.
Secondary keywords: deep learning;neural networks;QRS complex;computer science;computer and information science;diploma;
Type (COBISS): Bachelor thesis/paper
Study programme: 1000470
Embargo end date (OpenAIRE): 1970-01-01
Thesis comment: Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages: 60 str.
ID: 11222868