Dados do Trabalho

Título

Development and validation of a sleep staging classification model for healthy adults based on a single-channel EEG device

Introdução

Despite its recent expansion, the validation of wearable devices for sleep monitoring is still deficient. Data regarding sleep stages are inaccurate and do not reflect those obtained by polysomnography, the gold standard method in Sleep Medicine, such as results for "light sleep" and "deep sleep", stages that do not correspond to the ones formally defined by the American Academy of Sleep Medicine.

Objetivo

To perform the techno-scientific validation and evaluate the accuracy of a classification model for sleep staging using a single-channel electroencephalogram (EEG) device in association with wrist actigraphy for the development of a classification model for sleep staging.

Métodos

We included healthy individuals aged between 20 and 40 years, no sleep disorders. The procedure consisted of one full-night type-I polysomnography in which two sets of additional devices were tested: (A) flexible EEG band+actigraph and (B) rigid EEG band+actigraph. Data were segmented according to the polysomnography stages (30-second epochs), the specific characteristics of each sleep stage were extracted, and the classification model was developed. The model was tested to verify its final accuracy (sensitivity, specificity tests and an error matrix). The prior sample size calculation showed a minimum of 8 individuals per combination of devices.

Resultados

We recruited 23 individuals (50% men, mean age: 31 ± 4 years, mean BMI: 24 ± 3 kg/m2), 12 for combination A and 11 for B. We tested three classifiers (Decision Tree, Support Vector Machine, and Naive Bayes). For combination A, the mean accuracy was 94% and the individual accuracy for each stage was: 98% for wakefulness, 69% for N1, 97% for N2, 97% for N3 and 91% for REM. The highest error rate was observed for N1 (31%) - mostly confounded with REM (15%), N2 (8%), and wakefulness (8%). For combination B, the mean accuracy was 90% and the individual accuracy for each stage was: 83% for wakefulness, 78% for N1, 97% for N2, 99% for N3 and 96% for REM. The highest error rate was observed for N1 (21%) - mostly confounded with N2 (9%) and with wakefulness (13%). The actigraph presented a minor contribution to the classification model, maybe due to its reduced temporal resolution.

Conclusões

We developed a precise and accurate model for sleep staging for healthy adults on two single-channel EEG devices, with similar performance on both, which evidences the agnostic potential of the classification model.

Palavras -chave

Electroencephalogram; Medical software; Sleep tracker; Wearable devices