Can Patient-Specific Classification Improve the Accuracy of Sleep Apnea Detection From the ECG?

This study deals with detection of sleep apnea from the ECG and was motivated by the disproportional additional effort which seems necessary to increase the detection accuracy significantly above a level which is reachable by comparatively simple decision strategies. We suspected that inter and intra-individual variability caused by confounding factors might be the root of the problem and were interested in the improvement potential achievable by individually optimized classification thresholds and additional control for body-position. In 121 patients we performed 140 registrations of polysomnograms (PSGs) and time-synchronized Holter ECGs (sampling rate 1 kHz). The respiratory annotations of the PSGs were mapped onto consecutive, non-overlapping epochs of one minute duration and served as reference for apnea detection from the ECG. We extracted mean absolute QRS amplitudes from ECG lead I and used the energy of the modulation in the frequency band 0.018 Hz - 0.047 Hz as apnea-sensitive feature. Classification thresholds for three different approaches were determined: A - global threshold for the data set. B - individual threshold for each recording. C individual threshold for each body position within each recording. The body position information was extracted from the PSG. For each case the threshold value providing the greatest overall accuracy was selected. The global threshold yielded an overall accuracy of 75.5% for epochs of 1 min duration. Record-individual thresholds improved the accuracy to 84.5%, further advanced to 88.7% by separate thresholds for each body position within each record. The number of records with an individual accuracy below 70% decreased from 46 (case A) over 27 (case B) to 0 (case C). In conclusion, the gain in performance observed in our study highlights the significance of individual consideration of confounding factors and suggests a way to advance the detection accuracy of sleep-disordered breathing from the ECG.