Deep learning for automated sleep staging using instantaneous heart rate

Figures

• 

  Fig. 1. Neural network architecture for sleep stage classification. Instantaneous heart rate, resampled to 2 Hz and padded to 72000 samples (10 h) is used as input. 1200 overlapping patches containing 256 samples each are created, one for every 30 s epoch of the input. Convolutional layers are used to extract local features from each patch to an 128-node embedding layer. The local features are then concatenated to form a vector time-series of length 1200 and depth 128. Next, dilated convolutional layers are used to extract long range temporal features across the length of the vector. Finally, an output convolutional layer is used to output 1200 four-class probability distributions.

• 

  Fig. 2. Example of Inputs, labels and predictions. Hypnogram of a healthy subject shows typical 60–90 min sleep cycles with more deep sleep in the first half of the night and more REM in the second half. The predicted stages track the reference very well and are also “smooth” compared with the reference, i.e they have fewer rapid stage bouts that only last 1 or 2 epochs.

• 

  Fig. 3. Overall model performance. Confusion matrix for the SHHS dataset (a) normalized and (b) full counts. Confusion matrix for the CinC dataset (c) normalized and (d) full counts. (e) Dataset sizes and overall accuracy.

• 

  Fig. 4. Night by night accuracy distributions on SHHS and CinC datasets. a, e Histogram of all nights. The black lines mark the 10th, 25th, and the 50 percentile of nights. b, f Probability density distribution of nightly accuracy split by gender. c, g Probability density distribution of nightly accuracy split by age quartiles. d, h Probability density distributions of nightly accuracies split by presence of apnea.

• 

  Fig. 5. Distributions of sleep metrics derived from PSG reference and model predicted stages. The test population is split into groups by the clinical measures marked in the x-axis. The y-axis shows the values of sleep metrics—points represent the mean value of the metrics and the bars represent the standard deviations. The distributions calculated using the PSG reference stages and the algorithmically predicted stages show excellent concordance for fraction of REM during sleep and sleep efficiency. Deep sleep is underestimated by our model in favor of light sleep, yet the qualitative effect of clinical states on light and deep fractions is consistent between reference and predicted stages. a–d stage fractions and sleep efficiency with gender. N = 360, 440. e–h stage fractions and sleep efficiency with age. N = 164, 168, 198, 270. i–l stage fractions and sleep efficiency with apnea severity. N = 364, 314, 65, 57. m–p stage fractions and sleep efficiency with presence of hypertension. N = 436, 364. q mean REM bout lengths with apnea severity. The mean REM bout durations calculated from algorithmically predicted stages are higher than the PSG reference stages because the predicted stages…

• 

  Fig. 6. Sleep covariate regression analysis. The tables below show results of regression analysis with the covariates on each row as independent variables and our algorithm predicted sleep metrics on each column as the dependent variable. The metrics are adjusted for age and gender before regressing on apnea, and adjusted for age, gender and apnea before regressing on hypertension. Boxes where the covariate was a significant predictor are filled with values, the rest are filled with “NS”. Boxes colored in green are those where our results qualitatively agree with a previous study12 done using on PSG reference stages. Boxes colored in red are those where the previous study12 and our results disagree. N = 800. a Regression coefficients with 95% confidence intervals. b Percentage of variance of sleep measure explained by covariate calculated as 100 * R-squared along with p-values.

• 

  Fig. 7. More example hypnograms. a Hypnogram of an apneic subject (AHI = 70). Hypnogram shows similar 60–90 min sleep cycles but the cycles often contain significant wake periods. b Hypnogram with poor input data quality. Predicted hypnogram gets some of the structure correct but incorrectly classifies many deep and REM epochs.

• 

  Fig. 8. Demonstration on wrist-worn PPG. Example of a hypnogram predicted by our algorithm using heart rate extracted from wrist-worn PPG device compared with the hypnogram predicted by an EEG based automated sleep scoring device.