Outlier detection for training sets in an unsupervised functional classification framework: an application to ECG signals

A new method for unsupervised classification of multivariate functional data is proposed, and its application to ECG signals of PROMETEO (PROgetto sull’area Milanese Elettrocardiogrammi Teletrasferiti dall’ Extra Ospedaliero) project is performed and discussed. In fact, classification is a challanging task when data are curves, as well as outlier detection [1]. In particular, in the application to ECG signals it makes possible a semi automatic diagnosis of cardiovascular diseases, based only on ECG traces morphology, then not dependent on clinical evaluations. In [2], a real time procedure consisting of preliminary steps like reconstructing signals, wavelets denoising and removing biological variability in the signals through data registration is tuned and tested. Then, a multivariate functional k-means clustering of reconstructed and registered data is performed. Since the performances of classification method are to be validated through cross validation, it is mandatory a suitable training of the algorithm on data. Functional boxplots [4] and a multivariate extension of depth for functional data [3] are then used to detect outliers. The main focus of this work is then the application of these procedures to increase the algorithm predictive power robustifying selection criteria of data to be included in the training set, and then to perform an unsupervised multivariate functional classification of ECG signals, based on the sole ECG’s morphology.