Automated Cardiac Rhythm Diagnosis for Electrophysiological Studies, an Enhanced Classifier Approach

Bostock, J. (2014). Automated Cardiac Rhythm Diagnosis for Electrophysiological Studies, an Enhanced Classifier Approach. (Unpublished Doctoral thesis, City University London)

Text - Accepted Version
Download (6MB) | Preview


Heart function can be impaired by rhythm disturbances (cardiac arrhythmia), illustrated by electrocardiogram (ECG) recordings. Computerised arrhythmia diagnosis is well established for ECG’s but less for intracardiac electrophysiological (EP) testing. Accurate diagnosis is pre-requisite for delivering appropriate treatment to patients however existing algorithms misdiagnose a proportion of arrhythmias. Studies suggested artificial intelligence (AI) classifiers are accurate using ECG and intracardiac electrogram features and reviews suggested new features might augment diagnosis. This study aimed to develop an accurate cardiac rhythm diagnostic algorithm for electrophysiological (EP) studies with potential application as a generic rhythm classifier.
An ethically approved prospective clinical study collected clinical history, right atrial and right ventricular intracardiac electrograms, beat-to-beat cardiac stroke volume, body motion and body temperature data during EP studies. An iterative system development life-cycle was used, including knowledge management and classifier development sub-processes. Domain expert knowledge and clinical arrhythmia diagnosis were modelled, synthesised as AI classifiers and used to classify cardiac rhythms.
Data collected from 65 patients was pre-processed into instances for classifier inputs. Decision tree, naïve Bayes, neural network, support vector machine and inference engine classifiers developed using Matlab showed good performance and were combined as a production system in a mixture-of-experts multi-classifier system. 18 different rhythms were classified, with the naïve Bayes classifier used to classify 11 rhythms, decision tree 4 rhythms, neural network and support vector machine one each, unclassified instances by the inference engine classifier and final class allocation using decision rule. Production system showed overall correct clasification rate 0.960; error 0.040; mean sensitivity 0.855; mean specificity 0.977; mean κ 0.767; mean positive predictive value 0.792; mean negative predictive value 0.975; mean Pearson’s phi 0.787, with P < 0.004 (equivalent to P = 0.05 for 18 way Bonferroni comparison) supporting no difference with the gold standard. Correct classification, sensitivity, specificity, Cohen’s kappa and positive predictive value showed values of 1.0 for inappropriate sinus tachycardia, focal atrial tachycardia and ventricular tachycardia and > 0.9 for sinus node dysfunction and atrio-ventricular nodal/ junctional tachycardias. Temperature, accelerometry and QT interval were assessed as features by a comparison of algorithm performances with each feature removed and found not to affect classification performance. An evaluation showed 10 beat analysis performed better than 5 beat analysis.
Modelling of the clinical diagnosis process produced an AI based mixture-of-experts multi-classifier system, which accurately diagnosed different 18 cardiac rhythms. The naïve Bayes classifier performed best and classified 11 rhythms. Features for clinical symptoms and predisposing factors, atrial electrogram morphology and changes in stroke volume were found to influence rhythm classification. High performances encourage further development and potential future improvements include: a larger sample dataset; inclusion of His and coronary sinus electrograms; data mining for unknown features with significant influence on diagnosis; binary classification. The aim to classify rhythm using artificial intelligence suitable for use during EP studies was satisfied and the research hypothesis that it outperformed current algorithms was accepted. The system was likely to be able to accept updates but needs conversion as a precursor to use in a live clinical environment.

Item Type: Thesis (Doctoral)
Subjects: R Medicine
Divisions: School of Informatics > Centre for Health Informatics

Actions (login required)

View Item View Item


Downloads per month over past year

View more statistics