Work

Heart Murmur Detection with Phonocardiogram Recordings: Analysis of Ensemble Learning Model Performance within XAI Framework

Year: 2024
Type: article
Abstract: Accurate and reliable information on human heart health is key to its prognosis. Most recently, advanced machine learning and deep learning methods are aiding the doctors in decision making. However, it still evades them to understand how a ML or DL model is able to do so. This calls for use of ML/DL model performance interpretation frameworks to correlate a particular model’s performance with its internal architecture and functioning. In this study, an attempt is made to interpret the performance two different classification models that participated in the Physionet Challenge 2022. SHAP XAI framework-based interpretations of the performances of one heart murmur detection model and one clinical outcome prediction model are done. The heart murmur detection model selected for interpretation is a transformer-based deep neural network (T-DNN) whereas the clinical outcome prediction model selected for interpretation is a Random Forest boosted with AdaBoost boosting strategy. The dataset considered for model performance interpretations is the CirCor DigiScope dataset. The dataset contains phonocardiogram recordings, socio-demographic information and other auxiliary information. The T-DNN is trained on DWT features computed from segmented phonocardiogram signals for three-class (Present, Absent, and Unknown) classification task. The AdaBoost-RF is trained on collection of features including statistical measures, wavelet transform features, time-based and frequency-based features. ANOVA method is used to reduce the dimensionality of the total number of features to 110. The AdaBoost-RF performs a binary (Normal and Abnormal) classification task. The T-DNN model performed classification with overall accuracy of 90.23% whereas the AdaBoost-RF model performed classification with overall accuracy of 89.1%. Shapley importance plot, summary plot and Swarm charts are used to interpret the classification performance of the T-DNN and the AdaBoost-RF here. The study provides insights into the workings of advanced machine learning and deep learning models during detection and identification of heart health from phonocardiogram recordings.
Institution: Quantum University
Language: English
Field: Medicine
Open Access status: diamond
OpenAlex ID: W7217282627
FWCI:
Cites: 22
Cited by:
Related to: 10