Recognition And Quantification Of Sleep Apnea By Analysis Of Heart-Rate Variability Parameters
C. Maier, M. Bauch, H. Dickhaus
University of Heidelberg
Objectives: This study was performed within the scope of the Computers in Cardiology Challenge 2000. Our first aim was to discriminate patients suffering from sleep apnea from healthy persons. Moreover, we investigated the suitability of different heart rate variability (HRV) parameters for quantification of sleep apnea.
Methods: After upsampling of the ECG from originally 100 Hz to 1000 Hz by means of cubic spline interpolation and careful QRS-detection, RR-Intervals were extracted and served as a basis for further analysis. We investigated standard time domain HRV parameters (i. e. SDNN, pNN50, SDSD and RMSSD). Moreover, the fractional spectral radius (FSR) and an Entropy measure that both quantifiy the statistical signal complexity based on Eigenvector-Analysis of the lag-embedded RR-Series, were tested. All parameters were calculated on a minute-by-minute basis. Due to the small sample size of only 35 persons in the tests set we abstained in a first step from complex classificators and a high dimensional feature space, what could easily overestimate the classification results. So far, our classification procedure is based on simple thresholding.
Results for quantitative assessment of apnea: Although several reproducible patterns in the behavior of the standard HRV parameters as described above during and outside phases of apnea could be observed by visual inspection of the annotated training set, classification results were not very promising. The complexity measures, however, showed a surprisingly high degree of correlation with the annotations of the human expert. In a first attempt we achieved a result of 13073/17268 = 75.71% of correct classifications (entry 20000503.102146, entrant 9) on the tests set by simply thresholding the median filtered FSR-values of each minute.
Results for apnea screening: Both complexity measures also proved as excellently suited for apnea screening. Taking the median over all one-minute-values of FSR in each record resulted in 28/30 = 93.33% (entry 20000503.084214, entrant 9) of correct classifications.