Prediction and Classification of Apneic Events Using a Hybrid CNN-LSTM Network Applied with Statistical Tools

• Grant number: 25/19661-4
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: January 01, 2026
• End date: December 31, 2026
• Field of knowledge: Engineering - Biomedical Engineering
• Principal Investigator: Carlos Dias Maciel
• Grantee: Alexandre Moreira dos Santos Delfino
• Host Institution: Faculdade de Engenharia (FEG). Universidade Estadual Paulista (UNESP). Campus de Guaratinguetá. Guaratinguetá , SP, Brazil

Abstract

Sleep apnea is a prevalent disorder with serious cardiovascular and cognitive impacts. The standard diagnosis, based on manual analysis of polysomnography (PSG), is an expensive, time-consuming, and access-restricted process, which hinders large-scale screening. Existing artificial intelligence models for automated diagnosis often assume the synchronicity of physiological signals, ignoring the temporal delays between airflow obstruction (FLOW signal), respiratory effort (RIP signal), and the subsequent drop in oxygen saturation (SpO2). This gap can limit the accuracy and interpretability of the models. To address this problem, this project proposes an innovative methodology that integrates two stages: first, the quantification of these physiological delays using statistical tools, such as cross-correlation; second, the explicit use of these delays as features in a hybrid neural network (CNN-LSTM). Using a clinical database from FMRP-USP, it is expected that the proposed model, enriched with this temporal information, will demonstrate superior performance in detecting and classifying the different types of apnea (obstructive, central, and mixed). The result has the potential to lead to faster, more accessible, and more physiologically aligned diagnostic tools. (AU)