Theses and Dissertations

ORCID

https://orcid.org/0000-0001-9160-4608

Advisor

Wang, Haifeng

Committee Member

Marufuzzaman, Mohammad

Committee Member

Tian, Wenmeng

Committee Member

Ma, Junfeng

Date of Degree

5-15-2026

Original embargo terms

Visible MSU Only 6 months

Document Type

Dissertation - Campus Access Only

Major

Industrial & Systems Engineering

Degree Name

Doctor of Philosophy (Ph.D.)

College

James Worth Bagley College of Engineering

Department

Department of Industrial and Systems Engineering

Abstract

With the proliferation of biosensors and data acquisition technologies in the healthcare domain, artificial i ntelligence (AI) and machine l earning (ML) have gained s ignificant tr action in recent years. This research presents a comprehensive study of current trends in mental health analysis, with a specific focus on the application of machine learning techniques to multivariate, multi-channel, and multimodal biometric signals - particularly for sleep stage classification. Accurate sleep stage classification is vital for the diagnosis and treatment of sleep-related disorders. This dissertation proposes a novel hybrid deep learning model, CTB-NET, which combines Convolutional Neural Networks (CNNs), Transformer-based attention mechanisms, and Bidirectional Long Short-Term Memory (BiLSTM) networks to classify sleep stages using multichannel Polysomnography (PSG) data. The model is evaluated on a real-world dataset comprising PSG recordings from 50 patients obtained from a pediatric sleep center. These recordings include EEG, EOG, and EMG signals from healthy individuals. CTB-NET is compared against traditional machine learning algorithms (SVM, KNN, XGBoost) and deep learning models (CNN, LSTM, CONVLSTM, Transformer) using evaluation metrics such as per-class F1-score, overall accuracy, Cohen’s Kappa, and mean F1-score (MF1). The proposed architecture demonstrates superior performance, achieving an accuracy of 70%, a Kappa score of 0.54, and an MF1 of 0.54—outperforming all baseline models, especially in classifying challenging stages like N3 and REM. Hyperparameter tuning was conducted using KerasTuner with Bayesian Optimization, enhancing model convergence and generalization. While classification of the N1 stage remains difficult due to its transitional nature, CTB-NET exhibits robust and balanced classification, underscoring its potential as a state-of-the-art solution for automated sleep staging using PSG data.

Download

Share

COinS