Developing and Validating a Prediction Model For Death or Critical Illness in Hospitalized Adults, an Opportunity for...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support

Background:

• Hospital early warning systems increasingly use machine learning (ML) to predict clinical deterioration.
• The complementary role of ML predictions with physician and nurse judgment is not well understood.

Purpose of the Study:

• To train and validate an ML model for predicting patient deterioration.
• To compare ML model predictions with real-world physician and nurse predictions.

Main Methods:

• Developed and validated a neural network model using retrospective data (April 2011-April 2019).
• Compared ML predictions with prospective clinician predictions (nurses, residents, attending physicians) in a cohort study (April-August 2019).

Main Results:

• ML model achieved clinician-level accuracy for predicting in-hospital death and ICU admission (AUC 0.77 vs. 0.64).
• ML predictions were more accurate than clinicians for ICU admission.
• Combining human and ML predictions detected 49% of clinical deterioration events, improving sensitivity by 16% compared to clinicians alone.

Conclusions:

• ML models can effectively complement clinician judgment in predicting hospital patient deterioration.
• Human-computer collaboration offers significant potential for improving prognostication and personalized medicine.