Sleep awake detection from leg-worn wearables using deep sensor fusion.

The late-fusion CNN-BiLSTM model achieved the highest classification performance with an AUC of 90.94% in five-fold cross-validation.

• The late-fusion model outperformed both the early-fusion CNN-BiLSTM model and the traditional SVM baseline.
• Performance was evaluated using five-fold cross-validation on overnight recordings from 14 children.
• The area under the ROC curve was 90.94% for the late-fusion model.
• Two CNN-BiLSTM architectures were compared: one using early fusion and one using late fusion of raw multimodal inputs.

The study employed a leg-worn multimodal wearable device (RestEaze) that recorded PPG, accelerometer, gyroscope, and temperature signals for sleep-wake classification.

• RestEaze is described as a leg-worn multimodal wearable capturing photoplethysmography (PPG), motion from accelerometer and gyroscope, and temperature signals.
• Data were collected as overnight recordings from 14 children referred for ADHD evaluation.
• The device provides an alternative to wrist-worn devices, which 'often miss subtle movement or physiological changes.'
• Raw multimodal inputs were classified in short time windows to determine sleep or wake states.

A Support Vector Machine (SVM) using handcrafted features was implemented as a traditional baseline for comparison against the deep learning models.

• The SVM used handcrafted features derived from the multimodal wearable signals.
• This baseline was established to contextualize the performance of the CNN-BiLSTM deep learning models.
• Both deep learning models (early and late fusion) were compared against this SVM baseline.
• The deep learning models employed raw multimodal inputs rather than handcrafted features.

Clinically relevant sleep metrics including total sleep time, wake after sleep onset, sleep onset latency, and number of awakenings were derived from the model outputs.

• The derived metrics are directly relevant to sleep disturbances commonly observed in children with ADHD, including 'delayed sleep onset, shorter total sleep time, frequent awakenings, and daytime fatigue.'
• These metrics were computed from the sleep-wake classification outputs of the deep learning models.
• The ability to derive these metrics supports the clinical utility of the approach for pediatric ADHD evaluation.
• A temporal label-smoothing method was applied to further improve consistency of the derived metrics.

A temporal label-smoothing method was applied post-classification to improve the consistency of sleep-wake predictions over time.

• Label smoothing was applied as a post-processing step to the model outputs.
• The method is described as improving 'consistency' of the sleep-wake state classifications.
• This technique addresses potential frame-by-frame inconsistencies in the short-window classification approach.
• The smoothing step contributed to more reliable derivation of sleep metrics such as sleep onset latency and awakenings.

Children with ADHD are identified as a population with notable sleep disturbances that current monitoring tools inadequately address.

• Sleep disturbances in children with ADHD include 'delayed sleep onset, shorter total sleep time, frequent awakenings, and daytime fatigue.'
• Polysomnography is described as 'costly and complex,' and wrist devices 'often miss subtle movement or physiological changes.'
• The 14 study participants were children referred for ADHD evaluation, representing the target clinical population.
• The authors frame accurate detection of sleep issues as 'important for clinical care' in this population.