Deakin University
Browse

File(s) under permanent embargo

Binarized multi-gate mixture of Bayesian experts for cardiac syndrome X diagnosis: A clinician-in-the-loop scenario with a belief-uncertainty fusion paradigm

journal contribution
posted on 2023-05-29, 02:35 authored by Moloud AbdarMoloud Abdar, A Mehrzadi, M Goudarzi, F Masoudkabir, L Rundo, M Mamouei, E Sala, Abbas KhosraviAbbas Khosravi, V Makarenkov, UR Acharya, S Saadatagah, M Naderian, S García, N Sarrafzadegan, S Nahavandi
Cardiac Syndrome X (CSX) is a very dangerous cardiovascular disease characterized by angina-like chest discomfort and pain on effort despite normal epicardial coronary arteries at angiography. In this study, we used a CSX dataset from the coronary angiography registry of Tehran's Heart Center at Tehran University of Medical Sciences in Iran to develop several machine learning (ML) methods combined with uncertainty quantification of the obtained results. Uncertainty quantification plays a significant role in both traditional machine learning (ML) and deep learning (DL) studies allowing researchers to create trustable clinical detection systems. We propose a novel Mixture-of-Experts (MoE) model, called Binarized Multi-Gate Mixture of Bayesian Experts (MoBE), which is an effective ensemble technique for accurately classifying CSX data. The proposed binarized multi-gate model relies on a double quantified uncertainty strategy at the feature selection and decision making stages. First, we use a clinician-in-the-loop scenario with a belief-uncertainty paradigm at the feature selection stage. Second, we use Bayesian neural networks (BNNs) as experts in MoBE and Monte Carlo (MC) dropout for gates at the decision making uncertainty quantification stage. The proposed binarized multi-gate model reaches an accuracy of 85% when applied to our benchmark CSX dataset from Tehran's Heart Center.

History

Journal

Information Fusion

Volume

97

Article number

101813

Pagination

101813-101813

Location

Amsterdam, The Netherlands

ISSN

1566-2535

Language

en

Publisher

Elsevier BV