Non-invasive machine learning estimation of effort differentiates sleep-disordered breathing pathology

Area of Science:

• Sleep Medicine
• Respiratory Physiology
• Artificial Intelligence in Healthcare

Background:

• Obstructive sleep-disordered breathing (SDB) events are characterized by increased respiratory effort, traditionally measured invasively via esophageal pressure (P_es).
• Esophageal pressure monitoring, while the gold standard, is poorly tolerated by patients due to its invasive nature.
• Non-invasive polysomnography (PSG) measures are routinely collected but do not directly quantify respiratory effort.

Purpose of the Study:

• To investigate the application of machine learning to non-invasive PSG signals for accurate modeling of peak negative esophageal pressure (P_es).
• To develop a non-invasive method for quantifying respiratory effort in patients with sleep-disordered breathing.

Main Methods:

• Utilized a dataset of 1,119 patients with PSGs including P_es measurements.
• Employed non-invasive PSG signals: nasal pressure, oral airflow, thoracoabdominal effort, and snoring.
• Implemented a long short-term memory (LSTM) neural network for context-based mapping of non-invasive features to P_es values.
• Validated the algorithm prospectively using a hold-out dataset.

Main Results:

• The machine learning model demonstrated a median difference of 0.61 cmH₂O (IQR: 2.99 cmH₂O) between measured and predicted P_es.
• Model performance showed good correlation with actual P_es (ρ_median = 0.581, p=0.01).
• Predicted P_es significantly differed between normal breathing and obstructive SDB events, but not central apneas.

Conclusions:

• Machine learning can accurately predict peak negative P_es from non-invasive PSG signals.
• The developed system offers a non-invasive tool for quantifying respiratory effort, enhancing SDB characterization.
• This approach improves clinical practice by avoiding invasive procedures while providing crucial data for differentiating SDB types.