Sleep modelled as a continuous and dynamic process predicts healthy ageing better than traditional sleep scoring

Abstract

Background: In current clinical practice, sleep is manually scored in discrete stages of 30-s duration. We hypothesize that modelling sleep automatically as continuous and dynamic process predicts healthy ageing better than traditional scoring.

Methods: Sleep electroencephalography of 15 young healthy subjects (aged ≤40 years) was used to train the modelling method. Each 3-s sleep mini-epoch was modelled as a probabilistic combination of wakefulness, light and deep sleep. For 79 healthy sleepers (aged 20-77 years), 15 sleep features were derived from manual traditional scoring (manual features), 7 from the automatic modelling (automatic features) and 24 from a combination of automatic modelling with traditional scoring (combined features). Age was predicted with seven multiple linear regression models with i) manual, ii) automatic, iii) combined, iv) manual + automatic, v) manual + combined, vi) automatic + combined, and vii) manual + automatic + combined sleep features. Using the same seven sleep feature groups, two support vector machine and one random forest classifiers were used to discriminate younger (aged <47 years) from older subjects with fivefold cross-validation. Adjusted coefficients of determination (adj-R²) and average validation accuracy (ACC) were used to compare the linear models and the classifiers.

Results: The linear model and the classifiers using only manual features achieved the lowest values of adjusted coefficient of determination and classification validation accuracy (adj-R² = 0.295, ACC = 63.00% ± 16.22%) compared to the ones using automatic (adj-R² = 0.354, ACC = 65.83% ± 9.39%), combined (adj-R² = 0.321, ACC = 63.42% ± 8.78%), manual + automatic (adj-R² = 0.416, ACC = 67.00% ± 8.60%), manual + combined (adj-R² = 0.355, ACC = 72.17% ± 12.90%), automatic + combined (adj-R² = 0.448, ACC = 65.92% ± 7.97%), and manual + automatic + combined sleep features (adj-R² = 0.464, ACC = 70.92% ± 10.33%).

Conclusions: Continuous and dynamic sleep modelling captures healthy ageing better than traditional sleep scoring.

Keywords: Age; Gaussian mixture model; Machine learning; Polysomnography; Slow wave sleep; Spectral analysis.