Machine Learning Multimodal Model for Delirium Risk Stratification.

Area of Science:

• Clinical Informatics
• Artificial Intelligence in Healthcare
• Patient Safety

Background:

• Hospital delirium is a common complication, impacting patient outcomes.
• Automating delirium risk identification using machine learning (ML) can improve early intervention.
• Limited data exists on ML model performance for delirium risk stratification in real clinical settings.

Purpose of the Study:

• To develop, operationalize, and validate a multimodal ML model for delirium risk stratification in non-intensive care units.
• To assess the model's impact on clinical workflow and patient outcomes.
• To evaluate the model's performance in live clinical practice.

Main Methods:

• A quality improvement study using automated electronic medical records and natural language processing.
• Developed an ML model trained on data from patients aged 60+ admitted between 2016-2020.
• Validated the model in live clinical practice (March 2023-March 2024) and compared outcomes with a pre-ML cohort.

Main Results:

• The ML model achieved an area under the curve of 0.94 (95% CI, 0.93-0.95).
• Monthly delirium detection rates increased significantly from 4.42% (pre-ML) to 17.17% (post-ML) (P < .001).
• Post-ML cohort received lower daily doses of benzodiazepines and olanzapine compared to the pre-ML cohort.

Conclusions:

• A novel multimodal ML model can automate delirium risk stratification effectively in live clinical practice.
• The model demonstrates feasibility and acceptable performance, aiding delirium identification and care.
• This approach may optimize resource allocation for enhanced delirium management.