Sleep Classification With Artificial Synthetic Imaging Data Using Convolutional Neural Networks.

Area of Science:

• Biomedical Engineering
• Data Science
• Sleep Medicine

Background:

• Wearable devices collect multi-modal physiological data for health monitoring.
• Accurate sleep classification is crucial for diagnosing sleep disorders.
• Existing methods often struggle with complex physiological data from wearables.

Purpose of the Study:

• To introduce the Artificial Synthetic Imaging Data (ASID) Workflow for sleep classification.
• To evaluate the performance of the ASID Workflow using data from a wearable device.
• To compare the ASID Workflow against traditional machine learning algorithms.

Main Methods:

• The ASID Workflow creates synthetic images from Temporal E4 Data (TED) including heart rate, accelerometer, electrodermal activity, and skin temperature.
• A Convolutional Neural Network (CNN) was employed for supervised image classification of sleep periods.
• Performance was assessed across various data resolutions and heart rate inclusion scenarios, compared to logistic regression, SVM, random forest, k-NN, and LSTM.

Main Results:

• The ASID Workflow achieved a mean weighted accuracy of 94.7% across all tested settings.
• The ASID Workflow outperformed competing machine/deep learning algorithms, especially with low-resolution data and without heart rate.
• CNN demonstrated a lower computational cost per subject compared to other algorithms.

Conclusions:

• The ASID Workflow effectively utilizes multi-modal physiological data for enhanced sleep classification accuracy.
• Data resolution and heart rate modality significantly influence the workflow's performance.
• CNN's ability to leverage the topological structure of 2D images provides superior analysis of temporal and spatial dependencies.