Session S83.3

Automatic Detection Of Apnea With The Electrocardiogram

P. de Chazal, C. Heneghan, E. Sheridan,
R. Reilly, P. Nolan, M. O'Malley

University College - Dublin
Dublin, Ireland

Entry 20000503.171154, entrant 14. Score 13124/17268=76.00%. This study investigates accurate detection of epochs of clinically significant sleep apnea through automated scanning of surface lead ECG recordings, based on the Philipps-University database. This provides a "gold-standard" measure of the presence of apnea based on assessment of a clinical panel with access to additional measurements, with each minute of the recordings independently classified as normal or sleep apnea. The database contains signals from 70 subjects with approximately eight hours of data per subject. The ECG signal and classifications of 35 of the recordings are available for training and the other 35 classifications withheld for independent validation of classifiers. Unvalidated QRS detection points are provided for the ECG files. We also detected approximately 85% of P wave onsets. A criterion in developing our classifiers was to ensure they were insensitive to imperfectly detected QRS and P waves. An initial set of 37 features was created for apnea detection based on R-R and P-R intervals derived from each minute of data. The R-R features include mean and standard deviation of R-R rate, serial correlation coefficients and low-frequency spectral measures. We also included novel features based on Allan variance and detection of sequences of alternating bradycardia and tachycardia. The P-R features include mean and standard deviation of P-R intervals and serial correlation coefficients. Automatic feature selection from these 37 features based on minimisation of Wilk's lambda was considered. A branch-and-bound algorithm was used to find the optimal subsets of features. We evaluated several classifiers based on linear and quadratic discriminants, and feed-forward neural networks. To minimise the bias of the performance estimates of the classifiers 35-fold cross-validation was used. Each fold contained data from one subject only. The cross-validated results show that the best classifier resulted in a sensitivity of 61% and specificity of 82%, equivalent to an overall accuracy of 74%. When the classifier was independently validated on the test set it achieved an overall accuracy of 76%.