Methodological development study: Dynamic mask attention graph neural network for mechanical ventilation in elderly...

Area of Science:

• Critical Care Medicine
• Biomedical Informatics
• Geriatric Medicine

Background:

• The aging global population increases the demand for mechanical ventilation in geriatric intensive care unit (ICU) patients.
• Accurate intubation prediction is challenging due to complex physiology and sparse ICU data in elderly patients.
• Premature or delayed intubation in elderly ICU patients carries significant risks, including organ damage and increased mortality.

Purpose of the Study:

• To develop a precise intubation prediction model for elderly ICU patients.
• To address the challenges of complex physiological traits and data sparsity in geriatric critical care.
• To improve clinical decision-making regarding mechanical ventilation for older adults.

Main Methods:

• Retrospective analysis of ICU data from MIMIC-IV and eICU databases for patients aged over 65.
• Formulation of the intubation prediction task using a sliding window with strict temporal data splitting.
• Proposal of a dynamic mask attention graph neural network (DymaGNN) with dynamic heterogeneous graph structures and adaptive edge-weighting.

Main Results:

• DymaGNN achieved an AUC of 0.8363 (MIMIC-IV) and 0.8557 (eICU) for intubation prediction.
• The model maintained an AUC of 0.8115 even with a 15% data missing rate.
• Feature interaction graph visualization confirmed clinical knowledge, enhancing model interpretability and trust.

Conclusions:

• The proposed DymaGNN offers a robust method for mechanical ventilation prediction in elderly ICU patients.
• The model demonstrates high predictive accuracy and resilience to data missingness.
• Interpretable feature interaction graphs enhance trust and facilitate real-world clinical application for intubation decisions.