Reliable determination of sleep versus wake from heart rate variability using neural networks

Heart rate, heart rate variability (HRV), and sleep state are some of the common physiologic parameters used in studies of infants. HRV is easily derived from infant electrocardiograms (ECG), but sleep state scoring is a time consuming task using many physiological signals. We propose a technique to reliably determine sleep and wake using only the ECG. The method would be tested with simultaneous ECG and polysomnograph (PSG) determined sleep scores from the Collaborative Home Infant Monitoring Evaluation (CHIME) study. The advantages include high accuracy, simplicity of use, and low intrusiveness, with design including rejection to increase reliability valuable for determining sleep-wake states in highly sensitive groups such as infants. Learning vector quantization and multi-layer perceptron (MLP) neural networks are tested as the predictors. The manual PSG scored test set has 38,121 (67.8%) sleep and 18,076 (32.2%) wake epochs for a total of 56,197 epochs. The MLP classification of the entire test set resulted in 77.7% agreement with the PSG sleep epochs and 79.0% with wake. The rejection scheme applied to the MLP resulted in 28.9% of sleep and wake epochs meeting the rejection criterion. Of the remaining 39,946 epochs 86.0% are in agreement with the PSG sleep epochs and 85.4% with wake. After systematic rejection of difficult to classify segments, this model can achieve 85%-86% correct classification while rejecting only 30% of the data. This is an improvement of about 7.8% over a traditional model without rejection