Machine-learning-derived sleep-wake staging from around-the-ear electroencephalogram outperforms manual scoring and actigraphy

Abstract

Quantification of sleep is important for the diagnosis of sleep disorders and sleep research. However, the only widely accepted method to obtain sleep staging is by visual analysis of polysomnography (PSG), which is expensive and time consuming. Here, we investigate automated sleep scoring based on a low-cost, mobile electroencephalogram (EEG) platform consisting of a lightweight EEG amplifier combined with flex-printed cEEGrid electrodes placed around the ear, which can be implemented as a fully self-applicable sleep system. However, cEEGrid signals have different amplitude characteristics to normal scalp PSG signals, which might be challenging for visual scoring. Therefore, this study evaluates the potential of automatic scoring of cEEGrid signals using a machine learning classifier ("random forests") and compares its performance with manual scoring of standard PSG. In addition, the automatic scoring of cEEGrid signals is compared with manual annotation of the cEEGrid recording and with simultaneous actigraphy. Acceptable recordings were obtained in 15 healthy volunteers (aged 35 ± 14.3 years) during an extended nocturnal sleep opportunity, which induced disrupted sleep with a large inter-individual variation in sleep parameters. The results demonstrate that machine-learning-based scoring of around-the-ear EEG outperforms actigraphy with respect to sleep onset and total sleep time assessments. The automated scoring outperforms human scoring of cEEGrid by standard criteria. The accuracy of machine-learning-based automated scoring of cEEGrid sleep recordings compared with manual scoring of standard PSG was satisfactory. The findings show that cEEGrid recordings combined with machine-learning-based scoring holds promise for large-scale sleep studies.

Keywords: EEG; automated sleep scoring; ear EEG; home EEG; mobile EEG.

Conflict of interest statement

JKE, MABC, NS, VLR, SD, DJD and AS declare no conflicts of interest, including any involvement in organizations with a financial interest in the subject matter of the paper. KBM and MDV received a grant from Circadian Therapeutics to perform this study. MDV is a founding member of Circadian Therapeutics.

© 2018 The Authors. Journal of Sleep Research published by John Wiley & Sons Ltd on behalf of European Sleep Research Society.

Figures

Figure 1

Participant wearing both cEEGrid electrode array and polysomnography electrodes. Permission was obtained from the individual for the publication of this image

Figure 2

Comparison of electrode derivations. The metric is “CI” as defined in Equation (1). “TB” in this context means “top versus bottom” and FB(L) and FB(R) mean “front versus back”, for left and right ear, respectively. Values for TB(L) and TB(R) have been plotted as “detached” from the rest of the derivations because they have been positioned on the x‐axis outside of their place in the ordering (which would have hidden them inside the group of 56 “single‐electrode” derivations)

Figure 3

(a) Visualization of the three chosen derivations. FB(L) and FB(R) are within‐C, and L–R is the difference between the average within each C. (b) Analysis of reliability of the three chosen derivations. “Accepted data” is equivalent to gi in (1). FB(L) and FB(R) mean “front versus back”, for left and right ear, respectively; L–R, average of left electrodes versus average of right electrodes.

Figure 4

Representative hypnograms for a sleep recording. Top shows the manual scoring, using polysomnography data. Middle shows manual scoring using cEEGrid data and bottom shows automatic scoring after post‐processing, using leave‐one‐subject‐out cross‐validation

Figure 5

Comparison of classification performance for various automatic sleep staging algorithms and one manual classification. (a) Using only sleep–wake scores, (b) using five‐stage scoring. Labels on the x‐axis are described in detail in Table 2

Figure 6

Sleep statistics. See Table 2 for a description of the methods compared. For each plot four r2‐values can be calculated, according to how well the straight line fits to the scatter plot. The r2‐values for each plot are (in the order of the legend): total sleep (0.72; 0.84; 0.89; 0.61; 0.78), sleep efficiency (0.34; 0.49; 0.65; 0.25; 0.27), WASO (0.55; 0.70; 0.52; 0.42; 0.32), SOL (0.94; 0.96; 0.61; 0.67; 0.63), REM latency (0.04; 0.00; 0.03; 0.21; NaN). WASO, wake‐after‐sleep‐onset; REM, rapid eye movement; SOL, sleep onset latency

Figure 7

Analysis of classifier performance as a function of number of features. Box plots show results of random sampling from the feature space; solid line shows performance of N highest ranked features. We notice that the optimal number of features appears to be around 20–30, but also that the overfitting resulting from using the full set is relatively minor (given that the decrease in kappa is only from 0.58 to 0.53)