Morphological autoencoders for apnea detection in respiratory gating radiotherapy

Area of Science:

• Medical Physics
• Biomedical Engineering
• Machine Learning

Background:

• Respiratory gating training enhances patient proprioception to minimize organ motion during radiotherapy.
• Accurate classification of breathing patterns is crucial for effective respiratory gating.

Purpose of the Study:

• To develop an autoencoder-based system for multiclass classification of breathing patterns: regular, apnea, and unconstrained.
• To evaluate the system's performance on real-world and reference datasets.

Main Methods:

• Morphological analysis of respiratory signals using an autoencoder trained on regular breathing patterns.
• Utilizing the autoencoder's input-output correlation for classifier training and selection.
• Validation on a novel respiratory gating biofeedback training dataset and the Apnea-ECG dataset.

Main Results:

• Achieved high accuracies of 95 ± 3.5% and 87 ± 6.6% for breathing and apnea classification, respectively.
• Demonstrated the viability of a generalized model for characterizing diverse breathing patterns.

Conclusions:

• Autoencoders enable a data-driven approach to feature extraction for respiratory gating training, focusing on signal morphology.
• The proposed system is suitable for real-time implementation and potential transfer to other domains.