A balanced sleep/wakefulness classification method based on actigraphic data in adolescents.

Area of Science:

• Biomedical Engineering
• Sleep Science
• Artificial Intelligence

Background:

• Automated sleep-wakefulness classifiers using wrist actigraphy (ACT) data often exhibit class imbalance, overestimating sleep and underestimating wakefulness.
• This imbalance stems from optimizing overall accuracy rather than balanced sensitivity and specificity, with typical datasets over-representing sleep samples.

Purpose of the Study:

• To develop and evaluate a balanced Artificial Neural Network (ANN) classifier for improved sleep-wakefulness detection from minute-by-minute ACT data.
• To address the limitations of previous classifiers by incorporating a balanced dataset and novel features.

Main Methods:

• An Artificial Neural Network (ANN) classifier was developed using minute-by-minute wrist actigraphy (ACT) data.
• An 11-minute moving window was employed for data analysis, incorporating new features like time of day, median, and median absolute deviation.
• Sleep and wakefulness data were balanced to optimize ANN training, and results were validated against a polysomnogram-based hypnogram.

Main Results:

• The ANN classifier achieved an overall accuracy of 92.8%.
• High sensitivity (97.6%) for sleep detection was maintained, with significantly improved specificity (73.4%) for wakefulness detection.
• The geometric mean of sensitivity and specificity was 84.9%, indicating a well-balanced classification performance.

Conclusions:

• The developed ANN classifier demonstrates superior performance in automated sleep-wakefulness classification from ACT data.
• Balancing sleep and wakefulness data and incorporating advanced features significantly improves classifier accuracy and balance.
• This approach offers a more reliable tool for sleep monitoring using wrist actigraphy.