60. Creation and Comparison of a Machine Learning Decision Tree and Traditional Risk Score to Predict Ceftriaxone Resistance in Cancer Patients with E. coli Bacteremia

Abstract Background There are several clinical tools that have been developed to predict the likelihood of extended-spectrum β-lactamase producing Enterobacterales; however, the creation of these tools included few patients with cancer or otherwise immunosuppressed. The objectives of this retrospective cohort study were to develop a decision tree and traditional risk score to predict ceftriaxone resistance in cancer patients with Escherichia coli (E. coli) bacteremia as well as to compare the predictive accuracy between the tools. Methods Adults age ≥ 18 years old with E. coli bacteremia at The University of Texas MD Anderson Cancer Center from 1/2018 to 12/2019 were included. Isolates recovered within 1 week from the same patient were excluded. The decision tree was constructed using classification and regression tree analysis, with a minimum node size of 10. The risk score was created using a multivariable logistic regression model derived by using stepwise variable selection with backward elimination at level 0.2. The decision tree and risk score statistical metrics were compared. Results A total of 629 E. coli isolates were screened, of which 580 isolates met criteria. Ceftriaxone-resistant (CRO-R) E. coli accounted for 36% of isolates. The machine learning-derived decision tree included 5 predictors whereas the logistic regression-derived risk score included 7 predictors. The risk score cutoff point of ≥ 5 points demonstrated the most optimized overall classification accuracy. The positive predictive value of the decision tree was higher than that of the risk score (88% vs 74%, respectively), but the area under the receiver operating characteristic curve and model accuracy of the risk score was higher than that of the decision tree (0.85 vs 0.73 and 82% vs 74%, respectively). Figure 1. Clinical Decision Tree Table 1. Regression Model and Assigned Points for Clinical Risk Score Table 2. Statistical Metrics of Clinical Decision Tree and Clinical Risk Score Conclusion The decision tree and risk score can be used to determine the likelihood of whether a cancer patient with E. coli bacteremia has a CRO-R infection. In both clinical tools, the strongest predictor was a history of CRO-R E. coli colonization or infection in the last 6 months. The decision tree was more user-friendly, has fewer variables, and has a better positive predictive value in comparison to the risk score. However, the risk score has a significantly better discrimination and model accuracy than that of the decision tree. Disclosures Samuel L. Aitken, PharmD, MPH, BCIDP, Melinta Therapeutoics (Individual(s) Involved: Self): Consultant, Grant/Research Support


Creation and Comparison of a Machine Learning Decision Tree and Traditional Risk Score to Predict Ceftriaxone Resistance in Cancer Patients with E. coli Bacteremia
Background. There are several clinical tools that have been developed to predict the likelihood of extended-spectrum β-lactamase producing Enterobacterales; however, the creation of these tools included few patients with cancer or otherwise immunosuppressed. The objectives of this retrospective cohort study were to develop a decision tree and traditional risk score to predict ceftriaxone resistance in cancer patients with Escherichia coli (E. coli) bacteremia as well as to compare the predictive accuracy between the tools.
Methods. Adults age ≥ 18 years old with E. coli bacteremia at The University of Texas MD Anderson Cancer Center from 1/2018 to 12/2019 were included. Isolates recovered within 1 week from the same patient were excluded. The decision tree was constructed using classification and regression tree analysis, with a minimum node size of 10. The risk score was created using a multivariable logistic regression model derived by using stepwise variable selection with backward elimination at level 0.2. The decision tree and risk score statistical metrics were compared.
Results. A total of 629 E. coli isolates were screened, of which 580 isolates met criteria. Ceftriaxone-resistant (CRO-R) E. coli accounted for 36% of isolates. The machine learning-derived decision tree included 5 predictors whereas the logistic regression-derived risk score included 7 predictors. The risk score cutoff point of ≥ 5 points demonstrated the most optimized overall classification accuracy. The positive predictive value of the decision tree was higher than that of the risk score (88% vs 74%, respectively), but the area under the receiver operating characteristic curve and model accuracy of the risk score was higher than that of the decision tree (0.85 vs 0.73 and 82% vs 74%, respectively).  Conclusion. The decision tree and risk score can be used to determine the likelihood of whether a cancer patient with E. coli bacteremia has a CRO-R infection. In both clinical tools, the strongest predictor was a history of CRO-R E. coli colonization or infection in the last 6 months. The decision tree was more user-friendly, has fewer variables, and has a better positive predictive value in comparison to the risk score. However, the risk score has a significantly better discrimination and model accuracy than that of the decision tree.
Disclosures. Samuel L. Aitken, PharmD, MPH, BCIDP, Melinta Therapeutoics (Individual(s) Involved: Self): Consultant, Grant/Research Support Background. The optimal duration of antimicrobial therapy for uncomplicated Pseudomonas aeruginosa bloodstream infection (BSI) is unknown. We compared the outcomes of short and prolonged courses of antimicrobial therapy in adults with uncomplicated pseudomonal BSI.

Short-versus prolonged-courses of antimicrobial therapy for patients with uncomplicated Pseudomonas aeruginosa bloodstream infection
Methods. All patients with uncomplicated P. aeruginosa BSI admitted at a tertiary-care hospital from May 2016 to September 2020 were included. We compared the rate of recurrent P. aeruginosa infection and 30-day mortality among patients who underwent short (7-11 days) and prolonged (12-21 days) courses of antimicrobial therapy using propensity score analysis with the inverse probability of treatment weighting (IPTW) method.