Random Forest Models for Patient-Level Clinical Risk Prediction in Risk-Adjusted Hospital Outcome Reporting

• Activity Number: 635
• Type: Topic Contributed
• Date/Time: Thursday, August 8, 2013 : 8:30 AM to 10:20 AM
• Sponsor: Section for Statistical Programmers and Analysts
• Author(s): Milan C Seth*+ and Hitinder Gurm
• Companies: University of Michigan, BMC2 Clinical collaborative and University of Michigan
• Keywords: Random Forest ; Hospital Profiling ; Clinical Risk adjustment

Abstract

Background: We explored use of random forest models (RF) for case-mix adjustment in hospital outcome reporting as they are accurate, computationally efficient in large datasets, can accommodate large numbers of predictors and are robust to over-fitting. Methods: Patient baseline clinical characteristics were used in RF to estimate risk of contrast induced nephropathy (CIN) in Percutaneous Coronary Interventions (PCI). 15 variables with the greatest model determined importance were included in a reduced model. RF models were evaluated by AUC and NRI statistics. RF predicted risks were used in hierarchical mixed effects regression allowing hospital clusters. Random effect posterior modes were used to identify outlier hospitals. Results: Data for 68,573 PCIs from 1/2010 to 6/2012 was randomly divided into training (70%) and validation (30%) datasets. In validation data, RF had good discrimination (AUC: full 0.85, reduced .84,p< .01) and predicted hospital variability was largely explained by RF risk estimates. On-line decision support tool https://stage.bmc2.org/calculators/cin/ Conclusions: Random forests should be considered for case-mix adjustment in hospital profiling.