Electronic Prediction Rules for Methicillin-Resistant Staphylococcus aureus (MRSA) Colonization

Background: � Considerable resources have been dedicated to reducing the rate of MRSA infections.� One frequently-employed infection control tool is surveillance testing among patients not known to be MRSA colonized.� This process is costly, and false-positive tests lead to isolation of non-carriers.� This technique would improve if patients at high risk of colonization could be readily identified for testing using an automated system.

Objective: � To derive and validate automatable prediction rules for identifying patients at high risk for MRSA colonization at the time of hospital admission, and to compare them to previously published models.

Methods: � MRSA-colonization prediction rules were designed using only prospectively collected electronic data resident in a patient's electronic medical record (EMR) within a day of admission.� The strategies for developing the rules were as follows: 1) �Comprehensive' rule � any variables resident in the EMR; 2) �Parsimonious' rule � only variables remaining after those with P > 0.01 were removed in stepwise fashion from the �Comprehensive' model; 3) �Clinician Entry' rule � only variables readily available to an admitting physician after meeting a patient for the first time; 4) �Simple EMR' rule � only variables available to a rudimentary EMR based on registration, laboratory, and admission-discharge-transfer data; 5) �Centers for Disease Control and Prevention (CDC) Risk Factor' rule � only simple-to-obtain variables that have been strongly identified with healthcare-associated MRSA colonization in the past.� Rules were derived using multivariable modeling in a population of 23,314 patients consecutively admitted and tested for MRSA colonization at a US hospital, who did not have known MRSA colonization.� Rules were validated in a cohort of 26,650 patients admitted to two other hospitals.

Results: � MRSA admission prevalence was 2.2% and 4.0% in the derivation and validation cohorts respectively.� Multivariable modeling identified independent predictors of MRSA colonization among demographic, admission-related, pharmacological, laboratory, physiological and historical variables.� The five prediction rules varied in their performance, but all could be used to identify the 30% of patients who accounted for more than 60% of all MRSA colonization, and about 70% of all MRSA patient-days (see Figure for �Comprehensive' rule performance).� All rules could also identify the 20% of patients with a >8% chance of colonization, and the 40% of patients in whom colonization prevalence was 2% or less.� The new rules outperformed 2 published prediction rules (Figure).

Conclusions: � We report prediction rules that can be used to automate the triage of patients for MRSA admission testing.� The efficiencies introduced by these rules could result in substantial cost savings to infection control programs with potentially little sacrifice in program effectiveness.