| dc.description.abstract | Cardiovascular disease is a pressing problem worldwide, a majority of which can first present
themselves with similar symptoms to pulmonary pathologies, complicating the early diagnosis.
Unlike the modern diagnostic tools such as electrocardiogram (ECG), chest computed tomography
(CCT) or X-ray; auscultation is not only risk-free but also requires much simpler equipment, all
the while retaining the potential to provide just as much insight into the patients’ condition as its
competitors, especially when coupled with data processing. The problem then is the lack of
explainability due to the complexity and subjectivity of the audio signal. To tackle this, this thesis
attempted to build an artificial intelligence model based on convolutional neural network that can
predict the condition of the patient based on the lung sound recorded from their chest, and explain
that result with interpretation methods, in accordance with theoretical findings about lung sound,
pulmonary diseases and cardiovascular diseases. The techniques used were short-time Fourier
transform (STFT) as a feature extractor, principal component analysis (PCA) to further enhance
the input and Gradient-weighted Class Activation Mapping (Grad-CAM) to interpret the model.
The proposed model achieved a great result, with the best model achieving 95.36% of accuracy,
96.03% of precision, 96.80% of recall, and 96.41% of F1 score, which was outperforming all of
the sound-based models from other studies; and provided meaningful insight about the audio signal
with the use of the simple, intuitive STFT. By interpreting the model, observation showed that
signals in the range of 10-40 Hz was the most meaningful for predicting heart diseases, which
closely resembled physiology findings as this is almost identical to the occurrence range for heart
sound; and that the 0 Hz and 4 Hz are the two major frequencies that differentiate between patients
with lung disease and healthy patients. | en_US |
