Patient-Specific Predictive Modeling Using Random Forests: An Observational Study for the Critically Ill.

Area of Science:

• Computational biology
• Medical informatics
• Machine learning in healthcare

Background:

• Electronic health records enable personalized patient outcome prediction models.
• Identifying similar past patients is key to training personalized models.
• Previous work utilized cosine similarity for patient similarity metrics (PSM).

Purpose of the Study:

• Investigate the random forest (RF) proximity measure as a PSM.
• Evaluate RF PSM for personalized mortality prediction in intensive care unit (ICU) patients.

Main Methods:

• Utilized 17,152 ICU admissions from the MIMIC-II database.
• Extracted predictor variables from the first 24 hours of ICU stay.
• Trained patient-specific models using RF PSM, varying the number of similar patients (M).
• Compared RF PSM with logistic regression, decision trees, and RF models, including case-specific random forests (CSRFs).

Main Results:

• RF PSM demonstrated superior or comparable performance to cosine similarity PSM.
• RF and CSRF models achieved the highest predictive performance (AUCs: 0.839 and 0.832, respectively).
• Personalization using RF PSM did not improve performance for RF and CSRF models, but benefited others.

Conclusions:

• RF PSM facilitates good mortality prediction for several models.
• RF and CSRF models achieved top performance without personalization.
• RFs show potential for patient-specific outcome prediction, highlighting the importance of distinguishing predictor and similarity variables.