Automatic detection of Atrial Fibrillation

<p>Atrial Fibrillation is an abnormal arrhythmia of the heart and is a growingconcern in the health sector affecting 1% of the population. The incidenceof atrial fibrillation increases with age and has been found to be more detri-mental to long term cardiac health than previously thought. Sufferers arefive times more likely to experience a stroke than others. Often, atrial fib-rillation is asymptomatic and is frequently discovered only when a patient visits a hospital for other reasons. The detection of paroxysmal atrial fib-rillation can be difficult. Holter monitors are used to record the ECG overlong periods of time, but the resulting recording still needs to be analysed.This can be a time consuming task and one prone to errors. If a miniature,low-power, wearable device could be designed to detect and record whena heart experiences atrial fibrillation, then health professionals would havemore timely information to carry out better, more cost effective courses of treatment. This thesis presents progress towards development of such a device. Atrial fibrillation is characterised by random RR interval, missing Pwave and presence of atrial activity. The detection of the P wave and atrialactivity can be unreliable due to low signal levels and differences in wave-form morphology between subjects. The random RR interval appears tobe a more reliable method of detection. By analysing the ECG signal inboth the frequency and time domains, feature sets can be extracted for thedetection process. In this research, the Discrete Wavelet Transform is used to generate several sub-bands for analysing wave form morphology, and anumber of RR interval metrics are created for analysing the rhythm. All features are further processed and presented to a support vector machine classification stage for the ultimate detection of atrial fibrillation. Forty eight files from the MITDB database of the PhysioNet online ECG reposi-tory were downloaded and processed to form separate training and test-ing data sets. Overall classification accuracy for normal sinus rhythm was93% sensitivity and 95% specificity, and for atrial fibrillation, 95% sensitiv-ity and 93% specificity. These results were found to be sensitive to the ECG morphology of the individual subjects. This means that the system either needs to be trained on a greater number of ECG morphologies or perhaps trained on the morphology of the individual under investigation. Putting this issue aside, the research to date shows that it is reasonable to expect a small, low powered, wearable device, to be capable of automatically detecting whena heart experiences atrial fibrillation.</p>