A methodological comparison of risk scores versus decision trees for predicting drug-resistant infections: A case study using extended-spectrum beta-lactamase (ESBL) bacteremia.

BACKGROUND

Timely identification of multidrug-resistant gram-negative infections remains an epidemiological challenge. Statistical models for predicting drug resistance can offer utility where rapid diagnostics are unavailable or resource-impractical. Logistic regression-derived risk scores are common in the healthcare epidemiology literature. Machine learning-derived decision trees are an alternative approach for developing decision support tools. Our group previously reported on a decision tree for predicting ESBL bloodstream infections. Our objective in the current study was to develop a risk score from the same ESBL dataset to compare these 2 methods and to offer general guiding principles for using each approach.

METHODS

Using a dataset of 1,288 patients with Escherichia coli or Klebsiella spp bacteremia, we generated a risk score to predict the likelihood that a bacteremic patient was infected with an ESBL-producer. We evaluated discrimination (original and cross-validated models) using receiver operating characteristic curves and C statistics. We compared risk score and decision tree performance, and we reviewed their practical and methodological attributes.

RESULTS

In total, 194 patients (15%) were infected with ESBL-producing bacteremia. The clinical risk score included 14 variables, compared to the 5 decision-tree variables. The positive and negative predictive values of the risk score and decision tree were similar (>90%), but the C statistic of the risk score (0.87) was 10% higher.

CONCLUSIONS

A decision tree and risk score performed similarly for predicting ESBL infection. The decision tree was more user-friendly, with fewer variables for the end user, whereas the risk score offered higher discrimination and greater flexibility for adjusting sensitivity and specificity.