A Biomagnetic Field Mapping System for Detection of Heart Disease in a Clinical Environment

This PhD was inspired by the clinical demand for a system to triage chest pain and the untapped diagnostic potential of magnetocardiography (MCG) to reliably detect silent
ischaemic heart disease, which is responsible for the highest mortality rate of any single disease category. The aim was to develop a low cost and portable biomagnetic field mapping system capable of differentiating between healthy and diseased hearts within an unshielded hospital environment. This entailed the development of a system based on an array of magnetometers with sufficient sensitivity (104fT/ Hz at 10Hz) and noise rejection performance (68.4 ± 3.9 dB) to measure the small magnetic field associated with the heart beat, the magneocardiogram, within a much larger background noise.
The array of induction coil magnetometers (ICM) we developed had sufficient sensitivity and were robust to high amplitude noise. These sensors were also cheap to manufacture and capable of operating on battery power, allowing a low cost, portable device to be developed.
The key element that allowed us to achieve unshielded operation was the development of an algorithmic spatial filter, used as a substitute to operation within a magnetically shielded room. This coherent noise rejection (CNR) algorithm exploits the difference in spatial coherence between the local cardiac signals and the distant background noise sources. The observed coherence width during a clinical trial of the system within
a hospital ward was 2.8 ± 0.9 × 10 6 mm 2 . This allowed us to capture MCG signals with a signal to noise ratio of SNR QRS = 0.93 ± 4.43dB. The performance of CNR was found
to improve by 9dB per order of magnitude increase in environmental spatial coherence width. The coherence width can be increased by changes to hospital architecture, electromagnetic field regulation and device design optimisation.
The thesis also explores a variety of approaches to obtain binary diagnostic information from MCG, from traditional statistical learning on manually engineered features to
machine learning. I found that machine learning techniques, in particular convolutional neural networks (CNN), were able to capture more diagnostic information than traditional techniques and achieved world class prediction accuracy of 88% on the clinical trial dataset.