Deep learning classification of capnography waveforms: secondary analysis of the PRODIGY study

Area of Science:

• Medical Devices
• Artificial Intelligence
• Respiratory Monitoring

Background:

• Capnography monitors use CO2 thresholds for 'no breath' alarms.
• False alarms occur due to minor CO2 fluctuations or waveform artifacts.
• Accurate breath detection is crucial for patient safety.

Purpose of the Study:

• To evaluate a deep learning approach for classifying capnography waveform segments.
• To determine the accuracy of AI in distinguishing 'breath' from 'no breath' events.

Main Methods:

• A convolutional neural network was developed to classify 15-second capnography segments.
• Data from 400 participants across 9 North American sites (PRODIGY study) were analyzed.
• Internal-external validation was performed across hospitals.

Main Results:

• The neural network achieved 0.97 accuracy, 0.97 precision, and 0.96 recall.
• Performance remained consistent across different hospital sites during validation.
• The model demonstrated high reliability in classifying capnography waveforms.

Conclusions:

• Deep learning offers a promising solution to reduce false capnography alarms.
• Further research is needed to compare AI-driven alarms with standard methods.
• This AI approach could enhance respiratory monitoring safety.