Principal Investigator R Togneri Project g43

Centre of Intelligent Information Processing Systems, Machine VP

Department of Electrical and Electronic Engineering,

University of Western Australia

Co- Investigator Arkell I Rasiah

Centre of Intelligent Information Processing Systems, Department of Electrical and Electronic Engineering, University of Western Australia

Arrhythmia Detection in ECGs

A syntactic approach to detecting arrhythmias in ECGs has been developed at CIIPS. It was designed with the following considerations;

* as a generic methodology for recognising `shapes' within a one dimensional signal.

* recognition is based on the morphological characteristics of these `shapes' and the time interval between consecutive `shapes'.

* its application to a real world problem namely, the annotation of ECG traces.

The approach involves sliding a window along a signal where the signal within the window is modelled as a series expansion of N hermite function. An (N+1) dimensional vector is then formed from the N series expansion coefficients together with an estimated spread term. This vector is then passed to a PNN (Probabilistic Neural Network) to provide probability density function and posterior probability function estimates for each `shape' primitive. These likelihood estimates are then used, together with a timing matrix of inter- shape periods, by the post- processing stage to recognise the signal in the window as one of four `shape' primitives (ie. ectopic beats). As the window slides with each time sample, this process of recognition is repeated. On completion, the entire signal is annotated in terms of these shape primitives from which various cardiac arrhythmias, defined in terms of one or more of these primitives, are easily identified by a simple parsing algorithm.

What are the basic questions addressed ?

How suitable is a vector computer for executing the shape recognition code?

What are the results to date and the future of the work ?

Usage of the VP2200 has proved invaluable from two standpoints. Firstly the lengthy simulations of the initial model, with liberally chosen parameters, have facilitated the selection of a new set of conservative parameter values. This entailed a six- fold reduction in the sample rate and an almost 25% reduction in the vector size. All this was done in order that faster processing may be achieved while still preserving comparable annotation results to that of the initial model.

Secondly, the code for the ECG annotator when complied on the VP2200, generated an executable that `core dumped' in response to a memory leak. This bug was traced to an out- of- bound array call which was surprisingly not picked up by the executable of both the SPARC and SGI platforms. Interestingly enough, this bug did not affect the overall outcome of the annotation which made it `invisible' to all but the VP2200 platform. A significant portion of our service units was spent tracking down and removing this most subtle of bugs.

The processing times for a 30 second ECG record for the various platforms using the model for both liberal and conservative parameter choices are as follows:

Computing Platform Liberal Parameters Conservative Parameters

1) Sun SPARC 2 10500 seconds 600 seconds

2) SGI Indy 1250 seconds 130 seconds

3) SGI Onyx 670 seconds 50 seconds

4) VP2200 3600 seconds 110 seconds (~9% vectorisation)

To date, testing has revealed an average recognition rate of approx. 90% for the `liberal' model and preliminary testing of the `conservative' model has revealed comparable results.

What computational techniques are used and why is a supercomputer required ?

The computational techniques used by the ECG annotator are as follows :-

1) Preprocessing stage: removal of the linear trend component within the window by linear regression.

2) Feature extraction stage: estimation of the series expansion coefficients using pseudo- spectral transform techniques involving Gaussian quadrature methods. In addition, a spread term is estimated by gradient-descent optimisation methods based on LS estimation methods.

3) Vector quantisation stage: estimation of probability density function and posterior probability function using a Probabilistic Neural Network..

4) Post- processing stage: recognition of signal using a complex algorithm based on Bayes decision theory.

Since the above approach is repeated for each time sample of an ECG, the numerical expense is quite high especially when the model parameters such as sample rate, window size, vector size etc., are chosen liberally. Use of the VP2200 was sought to provide some insight into the algorithm's performance when generous parameter values are used. Given Arkell Rasiah has concluded the research component of his PhD and has commenced his thesis writeup, we will not be seeking an extension of our grant on ANU's VP2200. Also, we find CIIPS's newly acquired dual- processor SGI Onyx workstation is more than sufficient to meet our future testing needs. Nevertheless, our use of the VP2200 was instrumental in debugging and testing our ECG annotation algorithm. We are therefore most grateful to the Supercomputer Time Allocation Committee at ANU for allocating us time on the VP2200.


Syntactic Recognition of Common Cardiac Arrhythmias, A I Rasiah and Y Attikiouzel, 16th Annual International Conference of IEEE Engineering, Biology and Medicine Society, Baltimore, USA 3-6th November 1994.

A Syntactic Approach to the Recognition of Common Cardiac Arrhythmias within a Single Ambulatory ECG Trace, A I Rasiah and Y Attikiouzel, Australian Computer Journal , 102-112, 26 no 3, August 1994.

Note that since the above publications significant improvements to the post- processing stage have been made.