Time-series visual representations for sleep stages classification.

Area of Science:

• Biomedical Engineering
• Sleep Science
• Artificial Intelligence

Background:

• Polysomnography (PSG) is the gold standard for sleep stage classification but is expensive and impractical for home use.
• Smartwatches offer a convenient, non-invasive, and cost-effective alternative for continuous sleep monitoring.
• Traditional AI methods for smartwatch sleep analysis often use raw or extracted time-series data.

Purpose of the Study:

• To investigate the efficacy of transforming time-series smartwatch data into visual representations for improved sleep stage classification.
• To compare the performance of visual representation methods against traditional approaches for sleep stage classification using smartwatch data.

Main Methods:

• Time-series data from smartwatch accelerometer and heart rate sensors were converted into visual formats (Gramian Angular Field, Recurrence Plots, Markov Transition Field, spectrograms).
• Two-dimensional convolutional neural networks were applied to these visual representations.
• Image patching and ensemble methods were utilized to enhance classification performance.

Main Results:

• Gramian Angular Field, combined with patching and ensemble techniques, achieved the highest performance.
• The proposed method demonstrated superior accuracy, exceeding 82% for two-stage and 62% for three-stage sleep classification.
• This approach showed significant improvements over traditional methods, up to 8-9 percentage points.

Conclusions:

• Visual representations of smartwatch time-series data are highly effective for sleep stage classification.
• This method surpasses traditional approaches, offering a competitive and reliable alternative for sleep monitoring.
• The findings support the use of visual representations for enhanced health monitoring and timely interventions.