Development of Risk Score Model and Decision Tree Algorithm for Predicting Infections With CRKp in Colonized Patients

Overview

DETERMINE trial is a prospective multicenter multinational cohort study. This study will be carried out to predict the risk of bloodstream infections (BSIs) or other types of invasive infection with carbapenem resistant K.pneumoniae in patients being colonized by CRKp. The results of DETERMINE trial would be quite important to prevent unnecessary coverage of carbapenem resistant Klebsiella pneumoniae in empirical treatment of colonized patients. In this study, both risk score model and decision tree algorithm will be constructed and compared with each other in terms of sensitivity, specificity, positive predictive value and negative predictive value.

1\. Main Objectives: 1. Building a risk score model and decision tree algorithm with an acceptable certainty for early prediction of BSI or other invasive infections caused by CRKp in CRKp carriers. 2. Comparison of risk score model and decision tree algorithm in terms of sensitivity and specificity rates, positive predictive and negative predictive values 3. Evaluation of risk score model and decision tree algorithm in terms of convenience for routine clinical use 2\. Hypothesis 2.1 Main hypothesis: Presence of immunosuppression, invasive devices, gastrointestinal surgery within prior 3 months, multi-site colonization beside stool, older age, diabetes as a comorbidity, admission to ICUs, presence of carbapenemase gene in CRKp that causes colonization, type of carbapenemase (eg. blaOXA-48 and blaKPC ), high SOFA score, high APACHEII score, high ECOG score (\>2), short time interval between identification of colonization and development of BSI or other type of invasive infection are independent risk factors for development of subsequent BSI or other invasive infections in CRKp colonized participants 2.2 Secondary hypothesis: 1. Sensitivity and specificity rates are similar in risk score model and decision tree algorithm. 2. Positive and negative predictive values are similar in risk score model and decision tree algorithm. 3. Both risk score model and decision tree algorithm are feasible for usage in our daily practices. 3\. Study Design: Prospective multi-centre multinational study 4\. Setting and study period: Tertiary care hospitals form different parts of the world will be included in DETERMINE trial. Study period is scheduled to be 01.04.2020-01.04.2021 or longer until pre-defined sample size is attained. Each participant will be included only once at the time of first infectious episode (BSI or other type of invasive infections) caused by CRKp or any other bacteria even if more than one infectious episode is reported. Local (primary) investigators will collect microbiological and clinical data of participants and record these data in a standardized electronic case report form. Site investigators will screen all patients who are admitted to ICUs, burn ward, solid and bone marrow transplant units, all patients who are admitted to wards in which an outbreak with CRKp is detected, all patients who share the same room with a patient colonized or infected with CRKp by rectal swab with once a week interval regularly. The participants who have CRKp rectal colonization will be screened by axillary, inguinal and surgical wound screening as well. These patients will be followed for emergence of BSI or other types of invasive infection for 90 days starting from identification of rectal colonization (day 0). Periodic rectal screening with once a week interval will be applied for non-colonized patients during their hospitalization. Carbapenem resistance in Klebsiella pneumonia isolates will be determined according to 2015 CDC criteria. 5\. Sample Size: The sample size can be determined by the number of parameter included in logistic regression model. Ideally, the number of participants that should be involved is calculated by multiplying the number of parameters involved in multivariate logistic regression model with 40. Therefore, the sample size needs to be at least 520 if all hypothesized parameters take place in logistic regression model (Peduzzi P., Concato J., Kemper E., Holford T.R., Feinstein A.R. (1996), A Simulation Study of Number Per Variable In Logistic Regression Analysis, J.Clin.Epidemiology, Vol 12, 1373-1379). 6\. Follow-up The regular assessment of participants will be carried out during hospitalization and after discharge. it will be performed in once a month interval through 90-days follow-up. The participants will be primarily evaluated for occurrence of any type of invasive infection (blood stream or other type of invasive infections). The participants will be warned to apply to our centres until the completion of 90-day follow-up period when symptoms of infection develop (eg fever, chills). Also, participants will be informed to call physicians who are primary investigators of particular centre when they arrive to hospital clinic or emergency department. Primary investigators will evaluate participants and send their culture samples according to clinical necessity. After review of the results of cultures and clinical presentation, the patients will be allocated into one of three different groups constituted by (i) infections with carbapenem resistant Klebsiella pneumonia, (ii) invasive infections with other type of microorganisms and (iii) non-infectious groups. In statistical analysis, cases with bloodstream infection and other types of invasive infection will be grouped and analysed separately. If the participant is discharged before completion of the 90-days, he/she will be contacted in 30-days interval by telephone call to appraise the investigated outcomes. 7\. Microbiological Analysis: Antimicrobial susceptibility tests (AST) will be performed in each centre and not be repeated. Therefore, AST results will be obtained from hospital database. Multiplex PCR will be used for identification of type of carbapenemases among CRKp isolates (both isolates recovered in rectal swab sample and blood or other site cultures). EUCAST breakpoints will be applied for identification of carbapenem resistant Klebsiella pneumonia. Rectal swab processing and identification of CRKp colonization will be performed according to CDC recommendations as follows: Perirectal swabs will be placed into universals containing 5 ml tyriptic soy broth (Oxoid, UK) with an ertapenem 10 ug disk and send to the reference centre in each country for overnight incubation at 37 °C. The broth cultures are then inoculated onto MacConkey agar (Oxoid, UK), and lactose-fermenting colonies are identified with an automatized system such as API20E or Vitek-2 etc. and confirmed by MALDI-TOF Biotyper CA system (Bruker, Daltonics, Bremen, Germany). Carbapenem resistance will be determined phenotypically with ertapenem E-test (bioMerieux, France) and the isolates with a MIC value of \> 0.5 mg/L are reported as CRKp. In BSIs or other invasive infections, only the first CRKp isolate of each patient will be included. Multiplex PCR (Polymerase chain reaction) will be applied for identification of type of carbapenemase genes in clinical CRKp isolates. Clonal relatedness and sequence typing of CRKp isolates will be determined by using MLST (Multi-locus sequence typing) method. Both multiplex PCR and MLST will be performed in reference centres of each country. 8\. Statistical Analysis: In this study, the patients who develop BSI with CRKp vs. any bacteria vs. do not develop any infection will be matched in a 1:1:2 ratios by using relevant parameters. The same procedure will be applied for the patients who develop invasive infections without bacteremia. The variables are expressed as absolute numbers and their relative frequencies. Continuous variables are expressed as mean and SD if normally distributed, or as median and interquartile range (IQR) if non-normally distributed. Discrete variables for the matched pairs were compared by McNemar's test; for continuous variables we used the Wilcoxon test. All the variables being associated with CRKp BSI or other invasive infections in the crude analysis (p \<0.1) were included in a backward, stepwise multivariate logistic regression model, separately. Overall goodness of fit for model will be analyzed by Akiake's Information Criteria (AIC) and Nagelkerke's R-square. Discrimination of the model will be assessed by receiver-operator curve (ROC) characteristics. Internal validation of final risk score model and decision tree algorithm will be made with last 1/3 enrolled patients of all cohort (validation cohort). The data of first 2/3 recruited patients will be used to develop risk score model and decision tree algorithm (derivation cohort). The risk of development of subsequent BSI wtih CRKp vs. any other bacteria vs. patients not having any type of infections through 90-days follow-up in CRKp colonized patients will be assessed by cox proportional hazards models in separate analyzes as well as the risk of development of other invasive infections in the same groups through 90-days follow-up will be analyzed by the same statistical method. To develop the risk score, variables that maintain statistical significance in the multivariate regression model will be assigned a point value corresponding to the beta-coefficient divided by the lowest beta-coefficient identified in the regression model, and the resulting quotient will be multiplied by two and rounded to the nearest whole number. Summation of the points generated by the calculated risk factors will result in a quantitative score that will be assigned to each patient in the database. Optimal breakpoint will be assigned by the Youden's J statistic. In summary, a decision tree will be built using the following steps: (1) identification of the most appropriate single variable that can divide the dataset into 2 groups ("nodes"), best minimized impurity of CRKp BSI or other invasive infections with CRKp status in each daughter node, according to the Gini impurity criterion, (2) repetition of the branching process within each daughter node and subsequent generations of nodes and (3) cessation at "terminal" nodes when no additional variables achieve further reductions in node impurity by applying pre-defined cut-off values for branching.