Optimal intensive care outcome prediction over time using machine learning.

Area of Science:

• Critical Care Medicine
• Data Science in Healthcare
• Predictive Analytics

Background:

• Prognostication is crucial for risk adjustment and decision-making in the intensive care unit (ICU).
• Current prognostic models are limited, often relying on early data and linear models, neglecting temporal trends.
• Investigating dynamic prognostic changes during ICU stays is vital due to extended admission durations and time-limited therapy trials.

Purpose of the Study:

• To assess the predictability of ICU mortality using machine learning (ML) as a function of time.
• To evaluate the impact of incorporating temporal trend data into ML models for improved prognostic accuracy.
• To compare the performance of ML models against traditional methods like logistic regression and APACHE-II.

Main Methods:

• Utilized machine learning (ML) techniques to analyze physiological and demographic data over time during ICU admissions.
• Developed predictive models, including Deep Learning, to capture non-linear relationships and variable interactions.
• Incorporated trend data from physiological variables to enhance predictive capabilities.

Main Results:

• ML models significantly outperformed logistic regression and APACHE-II in predicting ICU mortality.
• Predictive power for mortality was maximal on the second day of ICU admission.
• Incorporating trend data further improved model performance, with the best ML model achieving an AUC of 0.895 on day two.

Conclusions:

• ML models incorporating time-series data offer superior risk stratification in the ICU compared to current tools.
• The peak predictive ability on day two supports the rationale for time-limited therapeutic trials.
• Dynamic prognostic assessment using ML and trend data can significantly enhance clinical decision-making in critical care.