1. 1 Significant pairwise comparisons based on a Bonferroni overall significance level of α = 0.10. 2 Parameter estimates for the interactions were examined by plots of parameter estimates for each combination of the two categorical variables. Background. Recent reports describing emerging non-susceptible (NS) Streptococcus pneumoniae (SP) isolates to levofloxacin (LEV) have become the focus of examination. N American SENTRY Program data (1997-2001) were analyzed to determine independent variables (ivs) predictive of MIC in hospitalized patients with SP blood isolates. Methods. MIC for LEV vs. patient-specific (e.g., age, specimen type, medical service category, infection risk factors) & hospital-specific (e.g., bed count, geographical (geo) region, study year (yr) ivs were analyzed using multivariable general linear modeling (GLM) for censored data with backwards stepwise elimination (at p > 0.1). Tree-based modeling was used to identify important ivs, breakpoints in continuous ivs, and possible interactions for inclusion into the GLM model. Results. LEV MIC 50, MIC 90, MIC range & % NS for isolates (n=383, from 28 hospitals) were: 1.0, 1.0,  0.25 to >4.0, 0.3. Significant ivs associated with percent increases in the geometric mean of MIC included: geo region (p<0.0001; Canada, Midwest, Southwest relative to Northeast; 17-23%), study yr (p<0.0001; other study yrs relative to 1997; 30-45%), and medical service category (p<0.012; acute care relative to medicine; 15%). Significant interactions were hospital duration by bed count (p=0.059) and age by primary diagnosis (p=0.007). This model explained 44% of the variance in MIC. When examining the performance of the model for predicting MIC for isolates within individual hospitals by study yr, observed vs. fitted mean MIC were highly correlated (weighted Spearman R = 0.92). Conclusions. Despite the narrow range in MIC distribution, a moderate proportion of variability in MIC was explained by these hospital- and patient- specific ivs. Identification of additional ivs, such as prior antibiotic exposure by isolate or institutional usage, would probably be needed to explain more of this variability. These data might be used to tailor empiric treatment and to prospectively curtail the growing incidence of NS SP to LEV and other fluoroquinolones. Relationships Between Susceptibility of Streptococcus pneumoniae Against Levofloxacin & Hospital- & Patient-Specific Variables: Report From the Antimicrobial Resistance Rate Epidemiology Study Team (ARREST Program) SM Bhavnani 1,2, JP Hammel 1, A Forrest 2, PG Ambrose 1, RN Jones 3 1 Cognigen Corporation, Buffalo, NY; 2 SUNY at Buffalo, NY; and 3 The JONES Group, North Liberty, IA 77 INTRODUCTION The Antimicrobial Resistance Rate Epidemiology Study Team (ARREST Program) was established as a collaborative effort between scientists from The JONES Group and Cognigen Corporation in order to use surveillance data and analytic techniques to better understand factors predictive of antimicrobial resistance. Data from surveillance programs such as the SENTRY Antimicrobial Surveillance Program have proven to be a valuable resource and have been used to alert the scientific and medical communities to problems of antimicrobial resistance. In response to the growing prevalence of antimicrobial-resistant S. pneumoniae, this analysis was undertaken to determine independent variables predictive of MIC in hospitalized patients with S. pneumoniae blood isolates. GLM Model Results Canada, Midwest, and Southwest U.S. were associated with 17-23% increases in mean MIC in comparison to the Northeast U.S. Study years after 1997 had a higher mean MIC than 1997 by 30-45%. Isolates from acute care patients had mean MIC values which were 15% higher than those from patients on medicine services. Significant two-way interactions included duration of hospital stay prior to pathogen isolation*hospital bed count, and primary diagnosis*patient age. The model explained a moderate proportion of the variability in MIC between patients (R 2 = 44%). However, when fit by institution as shown in Figure 2, model-predicted mean MIC within hospitals was strongly correlated with observed mean MIC averaged over all study years (Spearman correlation = 0.92). A high correlation was also seen when averages were computed for institution by year (Spearman correlation = 0.83). Data Collection Patient- and institution-specific and susceptibility data for S. pneumoniae blood isolates (one per patient) collected from North American hospitals participating in the SENTRY Antimicrobial Surveillance Program (1997-2001) were queried for analysis. Primary Outcome The primary outcome variable was the in vitro activity of levofloxacin against S. pneumoniae which was measured by the minimum inhibitory concentration (MIC). Observed values of MIC included left- and right-censored values of the form  0.5 (or  0.25 for 1997) and >4, respectively. A log 2 transformation of MIC was used to achieve approximate normal error distributions. Using NCCLS interpretive criteria, MIC values were classified as susceptible (  2), intermediate (4), and resistant (>4). Independent Variables Patient-specific variables included age, sex, specimen type, medical service category, infection risk factors, primary diagnosis, duration of hospital stay prior to pathogen isolation, nosocomial infection, and residence in an ICU. Additional independent variables included study year and institution- specific variables (hospital bed count, geographic region, and formulary demographics). Tree-Based Modeling Using S-Plus 6.0.1 for Unix, tree-based modeling was carried out to identify subgroups that manifested impressive differences in MIC using recursive partitioning. Potential two-way interactions between independent variables for inclusion in regression modeling were identified. Multivariable General Linear Modeling (GLM) for Censored Data Using SAS  8.2, GLM for censored data was carried out. Continuous independent variables were categorized into subgroups (using breakpoints to define interpretable subgroups of sufficient size) to account for potential nonlinear relationships. The model was constructed using backward stepwise elimination (p>0.10). The proportion of error variance explained by the model (denoted as R 2 ) was used to measure model precision. A Spearman correlation measure was used to assess the strength of association between model-predicted and observed MIC means within institutions, across all study years and within study years. METHODS 383 S. pneumoniae blood isolates from 28 hospitals were collected. 7 and 5 hospitals were located in the Midwest and Northeast regions of the U.S. All other regions had 4 hospitals each. The variability in observed MIC was narrow with an estimated standard deviation of 0.60 on the log 2 scale and a range of observed MIC values from  0.25 to >4 (Figure 1). There were 67 (17%) left-censored observations (with MIC values of  0.25 or  0.5) and one right-censored observation. Proportions of isolates by categories of each independent variable are summarized in Table 1. Based on the results for the tree-based model, two-way interactions were selected for consideration in the multivariable GLM. RESULTS Figure 1: MIC Histogram of All S. pneumoniae Blood Isolates (N=383) Table 2: Parameter Estimates from the GLM Model for Censored Data Table 1: Proportion of Isolates by Selected Independent Variables These data demonstrate a significant increase in the MIC of levofloxacin to S. pneumoniae over the study period. The use of GLM tailored censored data allowed for prediction of MIC and estimation of the likelihood of non-susceptibility based on patient- and institution-specific factors. Given the independent variables evaluated, a moderate proportion of variability in the MIC at the patient level was explained. As part on our on- going efforts, the impact of additional patient and institution-specific variables on model-precision will be evaluated, including prior antimicrobial usage. This type of model can be further applied to identify subgroups of patients at increased risk for infection arising from organisms with increased MIC. Prospective identification of such risk factors and appropriate intervention may then arrest further progression of increasing MIC values. CONCLUSIONS ABSTRACT As shown by the Line of Reference, the observed MIC distribution shifts to the right as the model-predicted MIC increases. Isolates in the lower third of the model-predicted MIC distribution were all susceptible with an extremely low estimated probability of non- susceptibility by a fitted normal curve. Isolates in the middle third were all susceptible with only a low probability of non-susceptibility, though the observed mean MIC was higher than that of the lower third. Only one of the isolates in the upper third was non-susceptible. The probability of non-susceptibility as estimated by a fitted normal curve was low. The percent of isolates from Canada, the Midwest, and the Southwest U.S., combined, represented 35%, 41%, and 60% of the lower, middle, and upper third subgroups based on the model-predicted MIC values. Specifically, these percentages for isolates from Canada were 18%, 14%, and 23%, respectively. Figure 2: Mean Model-Predicted MIC vs. Mean Observed MIC at the Institution Level Figure 3: Histograms of Observed MIC for Isolate Subgroups Falling in the Lower, Middle and Upper Third of Model-Predicted MIC Figure 3a: Lower Third (n=127) Figure 3b: Middle Third (n=128) Figure 3c: Upper Third (n=128) MIC (mg/L) Mean MIC = 0.68 MIC (mg/L) Line of Reference Mean Observed Log 2 MIC Mean Model-Predicted Log 2 MIC Censoring Boundary For more information, please contact: Sujata M. Bhavnani, Pharm.D. Cognigen Corporation 395 Youngs Road, Buffalo, NY, 14221 sujata.bhavnani@cognigencorp.com 716-633-3463, ext. 273