Advancing In-Hospital Clinical Deterioration Prediction Models

Area of Science:

• Medical informatics
• Clinical prediction models
• Cardiopulmonary resuscitation

Background:

• Early warning systems (EWS) for patient deterioration often lack robust evidence of improving outcomes.
• This limitation may stem from EWS primarily predicting binary outcomes rather than time-to-event predictions.

Purpose of the Study:

• To compare the predictive accuracy of logistic regression and random forest (classification models) against Cox proportional hazards regression and random survival forest (time-to-event models).
• The study focused on predicting in-hospital cardiopulmonary arrest.

Main Methods:

• A retrospective cohort study design was employed.
• Prediction models were developed using deidentified electronic health records from an urban academic medical center.

Main Results:

• Classification models (logistic regression, random forest) demonstrated comparable or superior statistical recall and precision to time-to-event models.
• Time-to-event models (Cox regression, random survival forest) provided predictions that could potentially offer better clinical guidance on the timing of cardiopulmonary arrest.

Conclusions:

• Refining early warning scoring systems necessitates the adoption of superior analytical methods.
• These methods should effectively model the phenomenon of patient deterioration while delivering understandable predictions for clinical use.