Deep learning-based detection of cessation of breathing in pre-term infants

Detecting apnoea of prematurity (AOP) in the Neonatal Intensive Care Unit (NICU) is challenging because current bedside monitors rely on simple thresholds that often produce high false-positive rates or miss short, irregular events. This study investigates whether deep learning architectures can improve the detection of Cessation Of BrEathing (COBE) events using routinely acquired physiological signals—specifically impedance pneumography (IP), electrocardiography (ECG), and photoplethysmography (PPG)—without requiring additional hardware.

The researchers compiled a dataset of 346 COBE and 608 non-COBE events from 430 hours of recordings across 24 pre-term infants. They evaluated three distinct deep learning architectures: a shallow convolutional neural network (CNN), various depths of residual networks (ResNets), and the ConvNeXt architecture. These models were tested using both individual signal modalities (unimodal) and combined inputs (multimodal fusion) to determine which configuration most effectively identified respiratory pauses.

The study found that the choice of input signal modality had a significantly greater impact on detection performance than the complexity of the neural network architecture. Models using IP signals consistently outperformed those relying on ECG-derived or PPG-derived respiratory surrogates. While multimodal fusion provided modest performance gains, the most effective configuration was a ConvNeXt model combining IP and PPG inputs, which reached a balanced accuracy of 88.7% and an F1 score of 0.75 on an independent test set.

These results demonstrate that sophisticated signal processing can enhance the reliability of neonatal monitoring using existing clinical infrastructure. By reducing the reliance on simple threshold-based alarms, such models could help mitigate alarm fatigue in NICUs and provide more accurate, timely identification of clinically significant respiratory events in vulnerable pre-term infants.