Decision Fusion in Automated Sleep Apnea Classification Using Multple Polysomnography Sensors and Convolutional Neural Networks

We propose a decision fusion approach using a multi-input Convolutional Neural Network (CNN) for sleep apnea detection. Four sensor modalities were analyzed using a nasal pressure transducer (NP), oronasal thermal airflow (TH), electrocardiogram (ECG), and oximeter (SpO2), signals from polysomnography (PSG) data from 130 patients attending a sleep clinic. The best-performing model, utilizing all four sensors, achieved over 94% accuracy, precision, specificity, and over 92% on recall. Combining the less obtrusive ECG with any airflow sensor resulted in improvements compared to ECG or airflow sensors alone, with accuracy reaching over 94
^%
. The results highlight the potential for decision fusion methods to improve sleep apnea detection by incorporating alternative sensors and leveraging the features automatically extracted by the CNN for each modality. This approach can optimize sensor usage for different monitoring environments, such as clinical settings or at- home monitoring, and help to find a balance between accuracy and unobtrusive monitoring.