Detection and classification of heart opening snaps

Damsgaard, Oliver Thomsen ; Bruun, Simon

4. term

Biomedical Engineering and Informatics, Master

2019

2019-06-06

46

Kardiovaskulære sygdomme er den største dødsårsag globalt, men diagnosticering bygger stadig ofte på lægens lytning med stetoskop (auskultation) og i nogle tilfælde invasive undersøgelser. For at mindske den subjektive vurdering afprøves nye computerbaserede metoder, der automatisk kan opdage hjertelidelser ud fra optagelser af hjertelyde (fonokardiografi, PCG). Unormale hjertelyde kan afspejle mekaniske problemer i hjertet. Et opening snap (OS) er et skarpt lydklik, som kan opstå når mitralklappen er stiv eller forsnævret (mitralstenose), ofte efterfulgt af en mislyd (en susende lyd fra turbulent blod). Dette projekt undersøger, hvorfor et OS nogle gange opstår uden en ledsagende mislyd, og om det kan hænge sammen med forkalkning i hjertet. Vi udviklede et system, der kombinerer parallelle Fully Convolutional Networks (FCN) og Long Short-Term Memory (LSTM) neurale netværk efterfulgt af en Support Vector Machine (SVM) klassifikator for at afgøre, om et OS er til stede i en PCG-optagelse. Tre parallelle netværk analyserer forskellige repræsentationer af den samme hjertelyd: (1) et filtreret råsignal, (2) Mel-Frequency Cepstral Coefficients (MFCC), en kompakt beskrivelse af lyde brugt bl.a. i talegenkendelse, og (3) features fra Discrete Wavelet Transform (DWT), som opfanger mønstre på forskellige tidsskalaer. I modsætning til mange studier analyserer vi kun den del af hjertecyklussen, hvor et OS forventes, ikke hele cyklussen. Denne målrettede tilgang fungerede godt med en bedste gennemsnitlig nøjagtighed på 92 % og et areal under kurven (AUC) på 0,9288. Derimod fandt vi ikke signifikante resultater for årsagen til OS uden ledsagende mislyd blandt de undersøgte faktorer, herunder en mulig sammenhæng med forkalkning.

Cardiovascular diseases are the leading cause of death worldwide, yet detection still often relies on a doctor listening with a stethoscope (auscultation) and, in some cases, invasive tests. To reduce subjective judgment, new computer-based methods are being tested to automatically detect heart disorders from recordings of heart sounds (phonocardiography, PCG). Abnormal heart sounds can reveal mechanical problems. An opening snap (OS) is a sharp sound that can occur when the mitral valve is stiff or narrowed (mitral stenosis), often followed by a murmur (a whooshing sound from turbulent blood). This project examines why an OS sometimes occurs without an accompanying murmur and whether this could be related to calcification in the heart. We built a system that combines parallel Fully Convolutional Networks (FCN) and Long Short-Term Memory (LSTM) neural networks, followed by a Support Vector Machine (SVM) classifier, to determine whether an OS is present in a PCG recording. Three parallel models analyze different representations of the same heartbeat: (1) a filtered raw signal, (2) Mel-Frequency Cepstral Coefficients (MFCC), a compact sound description used in speech processing, and (3) Discrete Wavelet Transform (DWT) features, which capture patterns at different time scales. Unlike many studies, we analyze only the part of the heart cycle where an OS would occur, not the entire cycle. This targeted approach performed well, achieving a best average accuracy of 92% and an area under the curve (AUC) of 0.9288. However, among the factors examined, we did not find significant evidence for the cause of OS without an accompanying murmur, including a potential link to calcification.

[This abstract was generated with the help of AI]

phonocardiography ; PCG ; cardiovascular disease ; CVD ; opening snap ; OS ; neural network ; long short-term memory ; fully convolutional network ; LSTM ; FCN ; FCN-LSTM ; detection ; classification

Download
View record in AAU Student Projects