Deep autoencoder-powered pattern identification of sleep disturbance using multi-site cross-sectional survey data

Area of Science:

• Computational Medicine
• Traditional East Asian Medicine (TEAM)
• Artificial Intelligence in Healthcare

Background:

• Pattern Identification (PI) is crucial for personalized Traditional East Asian Medicine (TEAM) treatments, including acupuncture and herbal medicine.
• Developing reproducible PI models using clinical data is essential for enhancing TEAM treatment effectiveness in real-world settings.
• Sleep disturbances are a common issue requiring effective diagnostic and treatment strategies.

Purpose of the Study:

• To propose a novel deep learning-based Pattern Identification (PI) model for sleep disturbance patients.
• To utilize a deep autoencoder for feature extraction and k-means clustering for patient stratification.
• To validate the model's performance using internal and external cluster validation metrics.

Main Methods:

• A cross-sectional study involving 2,579 sleep disturbance patients from the Republic of Korea Army (ROKA).
• Feature extraction using Principal Component Analysis (PCA) and a deep autoencoder, followed by k-means clustering.
• Internal validation (Calinski-Harabasz index, silhouette coefficient, within-cluster sum of squares) and external validation (PSQI, Berlin, GSRS, NQ scores).

Main Results:

• The deep autoencoder demonstrated superior performance in feature extraction for clustering compared to PCA and raw data.
• The model successfully identified three distinct Pattern Identification (PI) types based on sleep quality, dietary habits, and gastrointestinal symptoms.
• The optimal number of clusters was determined to be three using the elbow method.

Conclusions:

• The proposed deep learning model offers a reproducible method for Pattern Identification (PI) in Traditional East Asian Medicine (TEAM).
• This AI-driven approach can differentiate patient subgroups based on complex clinical data, aiding personalized treatment strategies.
• The model holds potential for developing AI-based clinical decision support systems and evaluating TEAM treatment efficacy.