Session P32.5

Developing an Alarm System for Smart Environments to Predict Paroxysmal Atrial Fibrillation via Neuro Fuzzy Modeling

B Pourbabaee*, C Lucas

Tehran University
Tehran, Iran

ECG shows heart activity and physicians try to diagnose heart disorders by analyzing ECG signals. Recently automated approaches are established to do tasks such as detecting different heart disorders, predicting onset of a heart attack and also classifying several arrhythmias. AF is one of the most common cardiac arrhythmia which causes irregular and repeated depolarization in atrium and hence this can have serious mortality and critical heart strokes. Therefore, it is useful to develop an alarm system which is able to gather patient ECG signal in order to predict PAF and reduce its latter effects such as thromboembolism. On the other hand, Neural Networks and Neuro-Fuzzy models are really powerful to predict chaotic time series. In this study, the Neuro-Fuzzy interpretation of locally linear models e.g. LOLIMOT is implemented to predict the exact time of PAF occurrence. Firstly, features such as mean, curve length, forth power, nonlinear energy, mean and energy of detail and approximation coefficients of wavelet transform, mixture density model parameters, P-wave width, RR interval and P-wave amplitude are extracted from ECG signal and then we select the best features which cause the most separation between three classes of far from PAF, preceding PAF and PAF by means of PCA and mutual information analysis. MI is more efficient than PCA. Now, we are able to use LOLIMOT model with incremental learning algorithm to predict the exact time of PAF. In this algorithm, the input space is divided in to small linear subspaces with fuzzy validity function. This is a fuzzy neuron and the process is continued until neuron enhancement doesn’t improve the performance and reduce the prediction error. Thus, this algorithm is enabled for long term prediction of exotic phenomena in smart environments. In order to increase the prediction horizon, sensitivity and decrease the false prediction rate we combine LOLIMOT with singular spectrum analysis to extract information from noisy chaotic time series, to provide an insight to the unknown dynamics and to enhance the signal to noise ratio. Finally the proposed method has been tested on PAF ECG signals from Physiobank for the Computers in Cardiology Challenge 2001.

(Abstract Control Number: 211)