Let Sleeping Patients Lie, avoiding unnecessary overnight vitals monitoring using a clinically based deep-learning...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Patient Monitoring

Background:

• Sleep impairment is prevalent in hospitalized patients, often due to frequent vital sign monitoring.
• Overnight monitoring disruptions are linked to adverse outcomes like delirium, cognitive decline, and increased mortality.
• Existing methods for reducing monitoring rely on subjective assessments or complex data, posing limitations.

Purpose of the Study:

• To develop and validate a predictive model for overnight patient stability.
• To enable the safe reduction of routine vital sign monitoring for hospitalized patients.
• To improve patient sleep quality and reduce associated healthcare burdens.

Main Methods:

• A recurrent deep neural network was trained on ~2.3 million admissions and 26 million vital sign assessments (2012-2019).
• The model utilizes sequences of five vital signs, a Modified Early Warning Score, and patient age.
• The algorithm is designed to be agnostic to patient location, condition, and demographics.

Main Results:

• The model achieved high predictive accuracy with an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.966 (retrospective) and 0.971 (prospective).
• It allows for the safe avoidance of overnight monitoring in approximately 50% of patient-nights.
• Misclassification of stable patients occurred at a very low rate (2 in 10,000 patient-nights).

Conclusions:

• A novel deep learning model accurately predicts overnight patient stability using routinely collected vital signs.
• This approach facilitates a significant reduction in unnecessary overnight monitoring, enhancing patient sleep.
• The model offers a practical, deployable solution for improving hospital patient care and well-being.