Supervised deep learning with vision transformer predicts delirium using limited lead EEG.

Area of Science:

• Critical Care Medicine
• Neuroscience
• Artificial Intelligence in Healthcare

Background:

• Delirium affects up to 80% of critically ill patients, leading to increased institutionalization, morbidity, and mortality.
• Current clinical detection rates for delirium are below 40%, despite using validated screening tools.
• Electroencephalography (EEG) is the gold standard for delirium detection but is too resource-intensive for routine monitoring.

Purpose of the Study:

• To evaluate the efficacy of a rapid-response, limited-lead EEG combined with supervised deep learning using a vision transformer model for predicting delirium.
• To assess the feasibility of this novel approach for delirium monitoring in mechanically ventilated, critically ill older adults.

Main Methods:

• Prospective proof-of-concept study involving mechanically ventilated, critically ill older adults.
• Utilized a rapid-response EEG device and supervised deep learning models, specifically a vision transformer architecture.
• Analyzed fifteen different models to predict delirium.

Main Results:

• Vision transformer models achieved 99.9%+ training accuracy and 97% testing accuracy when using all available data.
• The combination of rapid-response EEG and vision transformer demonstrated capability in predicting delirium.
• The monitoring approach was found to be feasible in the target patient population.

Conclusions:

• Rapid-response EEG coupled with vision transformer deep learning shows significant potential for accurate delirium prediction in critically ill older adults.
• This method could enhance delirium detection, enabling timely, individualized interventions.
• Potential benefits include reduced hospital length of stay, increased home discharge rates, decreased mortality, and lower healthcare costs.