Imputing missing sleep data from wearables with neural networks in real-world settings

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Sleep Medicine

Background:

• Accurate longitudinal sleep data is crucial for health but challenging to collect outside labs.
• Actigraphy provides rest-activity data but often has missing values, hindering analysis.
• Individuals with irregular sleep patterns pose unique challenges for sleep monitoring.

Purpose of the Study:

• To introduce SOMNI (Sleep data restOration using Machine learning and Non-negative matrix factorIzation), a novel neural network model for imputing missing actigraphy data.
• To evaluate the performance of SOMNI's individual and global approaches in handling missing sleep data, particularly for individuals with disturbed sleep-wake cycles.
• To provide clinicians with a tool for better managing missing sleep data and monitoring patients with irregular sleep patterns.

Main Methods:

• Developed a two-hidden-layer neural network model incorporating Non-negative matrix factorization (NMF).
• Implemented two data imputation approaches: individual (single-participant data) and global (multi-participant data).
• Validated the SOMNI model using rest-activity data from shift and non-shift workers across three hospitals.

Main Results:

• Both individual and global SOMNI approaches accurately imputed missing data for long datasets (>50 days), even for shift workers (AUC > 0.86).
• For short datasets (~15 days), only the global approach demonstrated accuracy (AUC > 0.77).
• The model effectively captured hidden longitudinal sleep-wake patterns in individuals with disturbed sleep.

Conclusions:

• SOMNI provides an accurate method for imputing missing actigraphy data, enhancing the monitoring of sleep-wake cycles.
• The model is particularly effective for long-term monitoring and individuals with highly irregular sleep patterns.
• This tool can improve clinical management of sleep disorders by enabling reliable data analysis outside laboratory settings.