Hannah Alphs Jackson, John Cashy, Ophir Frieder and Anthony J. Schaeffer*,

Transcription

1 Infection/Inflammation Data Mining Derived Treatment Algorithms From the Electronic Medical Record Improve Theoretical Empirical Therapy for Outpatient Urinary Tract Infections Hannah Alphs Jackson, John Cashy, Ophir Frieder and Anthony J. Schaeffer*, From the Department of Urology, Northwestern University, Feinberg School of Medicine, Chicago, Illinois (HAJ, JC, AJS), and Department of Computer Science, Georgetown University, Washington, D.C. (OF) Purpose: We determined whether data mining derived algorithms from electronic databases can improve empirical antimicrobial therapy in outpatients with a urinary tract infection. Materials and Methods: The electronic medical records from 3,308 visits associated with a positive urine culture at Northwestern s outpatient Urology and Internal Medicine clinics and Emergency Department from 2005 to 2009 were interrogated. Bacterial species and susceptibility rates for trimethoprim-sulfamethoxazole, ciprofloxacin and nitrofurantoin were compared. Using data mining techniques we created algorithms for empirical therapy of urinary tract infections and compared the theoretical outcomes from data mining derived therapy to those from conventional therapy. Results: Patients were significantly older in the Department of Urology vs Internal Medicine vs Emergency Department, and more patients in the Department of Urology were male. During the 5-year period the susceptibility rates for ciprofloxacin in the Department of Urology and trimethoprim-sulfamethoxazole in Internal Medicine decreased significantly. In the Department of Urology the susceptibility rate for nitrofurantoin was greater than for ciprofloxacin, which was greater than for trimethoprim-sulfamethoxazole. In all departments, bacteria were more resistant to trimethoprim-sulfamethoxazole than to ciprofloxacin or nitrofurantoin. All drugs were more effective in the Emergency Department and Internal Medicine than the Department of Urology. Prior resistance patterns were the strongest predictor of current susceptibility profiles. In the Department of Urology the algorithms for patients with or without prior cultures theoretically outperformed conventional therapy in men (13.2%) and women (10.1%). Conclusions: Antimicrobial resistance patterns in outpatient urinary tract infections are time dependent, and drug and site specific. Data mining directed therapy significantly improved theoretical outcomes compared to conventional therapy for Department of Urology outpatients and for female patients in the Emergency Department. Key Words: urinary tract infections, bacteria, data mining, anti-bacterial agents, drug resistance Abbreviations and Acronyms CIP ciprofloxacin ED Emergency Department EDW Enterprise Data Warehouse IDSA Infectious Disease Society of America IM Internal Medicine NFN nitrofurantoin NU Northwestern University TMP-SMX trimethoprimsulfamethoxazole URO Department of Urology UTI urinary tract infection Submitted for publication April 15, Study received institutional review board approval. Supplementary material for this article can be obtained at edu/depts/urology/faculty/academic/supplementals. * Correspondence: Department of Urology, Tarry , Northwestern University, 303 E. Chicago Ave., Chicago, Illinois ( ajschaeffer@northwestern.edu). Financial interest and/or other relationship with Advanstar Communications Inc., BMJ Group, UpToDate Inc., FlashPointMedica, Pinnacle Pharmaceuticals Inc. and Baylor Health Care System. See Editorial on page URINARY tract infections afflict millions of individuals in the United States causing significant morbidity and costs up to $3.5 billion annually. UTIs account for more than 9 million outpatient visits to physician offices annually. The lifetime risk of UTIs in women and men is approximately /11/ /0 Vol. 186, , December 2011 THE JOURNAL OF UROLOGY Printed in U.S.A by AMERICAN UROLOGICAL ASSOCIATION EDUCATION AND RESEARCH, INC. DOI: /j.juro

2 2258 DATA MINING FOR IMPROVED EFFICACY IN URINARY TRACT INFECTION THERAPY 60% and 12%, respectively. 1 3 In ambulatory women the national cost of uncomplicated UTIs approaches $2.5 billion per year, while costs of recurrent UTIs in this population are close to $1.6 billion. 4 7 In men the total annual health care expenditures for UTIs have also been estimated to be greater than $1 billion. 8 At a time when control of spiraling health care costs is a national priority, improving the management of UTIs is imperative. A presumptive diagnosis of a UTI is made by a history of irritative voiding symptoms and a urinalysis that suggests an infection but does not identify the bacterial species or antimicrobial susceptibility. 6 UTIs are difficult to treat because the pathogens and their related antimicrobial susceptibility vary widely depending on the clinical scenario. Because the bacterial pathogens responsible for uncomplicated UTIs in women are reasonably predictable, the IDSA recommends that urine cultures, which take up to 72 hours, not be performed. 9 In practice, even if a culture is done, UTI symptoms warrant empirical treatment driven by historical data of probable pathogens, antimicrobial susceptibility patterns and treatment guidelines. The 1999 IDSA guidelines for uncomplicated UTIs in women recommend 1 tablet of TMP-SMX double strength taken orally twice a day for 3 days. 2 When TMP-SMX resistance exceeds 10% to 20% in the community, empirical treatment of UTI should be switched to another agent, of which the fluoroquinolones (eg CIP) have the greatest efficacy. 2 To our knowledge no guidelines exist for empirical outpatient UTI therapy in men. However, as UTIs in men are considered complicated, a fluoroquinolone is often used as first line therapy. However, bacterial resistance in UTIs is increasing, and is site and patient sensitive. 10 Thus, reevaluation of these recommendations and practices appears warranted. We hypothesize that the uncertainty about the proper course of treatment for UTIs can be reduced and outcomes improved if comprehensive, up-todate data are available to generate and maintain patient and site specific UTI management guidelines. MATERIALS AND METHODS Northwestern Memorial Hospital is an 873-bed teaching hospital for Northwestern University s Feinberg School of Medicine. Physicians at Northwestern s Urology, Internal Medicine and Emergency Departments are affiliated with the Northwestern Medical Faculty Foundation. The Northwestern University EDW is a data warehouse containing electronic medical records, laboratory results and billing data from Northwestern Memorial Hospital and Northwestern Medical Faculty Foundation for patients treated in the last dozen years. We queried the NU EDW to obtain the results of urine cultures for patients who were seen as outpatients between January 1, 2005 and December 31, 2009 with a diagnosis of UTI, symptoms of UTI (ie dysuria) and a positive urine culture. In addition to ICD-9 coding (associated with the office visit or the urine culture order), ED, IM and URO visits were queried for the phrases UTI, urinary tract infection or dysuria. A positive urine culture was defined as a uropathogen present in clinically significant colony counts (greater than 10,000 cfu/ml) for which susceptibility data were available. Results of the urine culture were linked to patient medical records, from which we obtained age, race, gender, zip code of billing address and other demographic characteristics for use in our analysis. Urine cultures contained the name of the organism detected and susceptibility to a laboratory determined panel of antimicrobials. We limited our analysis to NFN, CIP and TMP-SMX, the 3 commonly used antimicrobials at NU that can be administered orally for the treatment of UTIs. Susceptibility to each antimicrobial was considered a binary outcome (susceptible or resistant), with partial or intermediate susceptibility being categorized as resistant. If multiple organisms were detected, a drug was presumed to be effective if all of the bacterial species identified in the culture were susceptible. The NU EDW was also used to query records for prior urine cultures. Patients whose most recent culture was susceptible were categorized as not previously resistant. Patients whose most recent cultures were resistant were categorized as previously resistant. To evaluate the potential efficacy of data mining to improve empirical therapy for UTIs, we used an association rule algorithm to interrogate the data and derive algorithms for empirical therapy. Algorithms were derived for patients with (NU Algorithm ) and without (NU Algorithm ) previous cultures. We compared theoretical outcomes with data mining directed therapy to those with conventional therapy. For conventional empirical therapy in women we used the 1999 IDSA guidelines for uncomplicated UTI (ie TMP-SMX if resistance to TMP-SMX during the prior year in question was less than 20%, or CIP if resistance to TMP-SMX during the prior year was greater than 20%). 9 The 1999 guidelines were used because the 2010 IDSA guidelines were not available during the period queried. For women seen in the ED in the years 2005, 2006 and 2009, IDSA guidelines would have resulted in the use of TMP-SMX. For all other years and sites CIP would have been indicated because the resistance to TMP-SMX was greater than 20%. For conventional empirical therapy in men we assigned CIP in all cases. To determine theoretical outcomes for UTI therapies we made several assumptions. We assumed that each patient received the assigned antimicrobial. In addition, we assumed that if the bacterial pathogens were susceptible to the assigned antimicrobial, the treatment would have been effective (ie resulting in cure and resolution of symptoms). As a corollary, if the pathogens were resistant, the assigned treatment would have been ineffective. Statistical analysis was performed using SAS 9.2 and SPSS 19. A significance level of 0.05 was used

3 DATA MINING FOR IMPROVED EFFICACY IN URINARY TRACT INFECTION THERAPY 2259 Table 1. Demographic characteristics of patients presenting with outpatient UTI ED IM URO Sample size ,233 Mean SD age* % Gender: M F % Race: White Black Other or unknown * Age in ED, IM and URO all statistically significantly different. Male gender in URO statistically significantly different from that in ED and IM. Female gender in URO statistically significantly different from that in ED and IM. White race in IM statistically significantly different from that in ED and URO. Other or unknown race in IM statistically significantly different from that in ED and URO. throughout the study. This study was approved by NU institutional review board protocol #STU RESULTS Mean age, gender and race of the patients associated with cultures obtained from the ED, IM and URO are shown in table 1. Patients in the ED were significantly younger while the URO had a higher percentage of male patients. Uropathogen Prevalence The distribution of uropathogens was evaluated across departments. Escherichia coli was the most common pathogen at all sites, but accounted for a significantly smaller proportion of uropathogens in the URO compared to IM or ED. Conversely Enterococcus species and Pseudomonas aeruginosa were significantly more prevalent, and accounted for a significantly larger proportion of uropathogens in the URO than in IM or ED ( 0.05). In IM there were 12 (2.5%) cultures with more than 1 uropathogen compared to 16 (2.7%) cultures in the ED and 152 (6.5%) cultures in the URO. Prior cultures were available for 36.2%, 20.3% and 5.1% of the patients in the URO, IM and ED, respectively. Antimicrobial Susceptibility Susceptibility rates for TMP-SMX, CIP and NFN between and within departments from 2005 to 2009 are shown in figure 1. For all drugs the susceptibility rates were significantly lower in the URO than in the ED and IM. During the 5-year period in the URO, bacterial susceptibility to CIP decreased significantly from 69.3% in 2005 to less Figure 1. Comparison of individual antimicrobial susceptibility patterns among departments (A) and of antimicrobial susceptibility patterns in each department (B). Statistical comparisons were made among average values of all 5 years studied. Asterisk and bullet denote statistical significance at 0.05 level. Dagger and symbol denote statistical significance to p 0.05 level. Asterisk indicates susceptibility rates for all drugs were significantly lower in URO than in ED and IM. Dagger indicates susceptibility rates for TMP-SMX were significantly reduced over time in IM while those for CIP were significantly reduced over time in URO. Bullet indicates susceptibility rates for TMP-SMX were significantly lower than for CIP and NFN in all departments. indicates that in URO, susceptibility rate for CIP was also significantly less than for NFN.

4 2260 DATA MINING FOR IMPROVED EFFICACY IN URINARY TRACT INFECTION THERAPY than 60% in 2007 to 2009, while susceptibility to TMP-SMX decreased in IM from 85% to 68% (fig. 1, A). For the other drugs and departments the bacterial susceptibility patterns remained stable. In all departments bacteria were significantly more resistant to TMP-SMX than to CIP or NFN, while in URO the bacteria were also more resistant to CIP than to NFN (fig. 1, B). Data Mining Demographic, clinical, microbiological and pharmacological data were interrogated with data mining techniques to determine the characteristics that could predict antimicrobial susceptibility of urine cultures and guide empirical therapy. Because of the susceptibility differences among sites, the data from each site were mined independently. At all of the sites the most effective antimicrobial in the prior year was highly predictive of the most effective drug for current therapy. Because NFN performed well in all departments and over time, we used various logistic regression models using age, race, gender and prior resistance patterns to help predict resistance to NFN. We found that gender and race were significant predictors of resistance, with men and African-American patients more likely to be resistant. However, even among these groups of patients NFN was superior to the other antimicrobials. Attempts to use logistic regression models to identify those patients who would respond better to TMP-SMX or CIP than to NFN were unsuccessful. Of the URO encounters 36% had prior cultures. In a logistic regression model prior resistance patterns were the strongest predictor of current susceptibility profiles, with more than 3 times increased odds of having a UTI resistant to NFN if a previous culture from the same patient grew pathogens resistant to NFN (OR 3.4, 95% Wald confidence limits ). The time between the past and current culture did not affect the predictive value of prior cultures. Identical analyses in IM and ED were not statistically significant. Data Mining Derived Algorithm We asked whether information from data mining could help improve empirical therapy choices for outpatient UTIs. Using data from each department we created 2 separate algorithms for patients with (NU Algorithm ) and without (NU Algorithm ) prior urine cultures. For NU Algorithm, the antimicrobial with superior susceptibility from the prior year was recommended. For NU Algorithm, the most effective drug from the prior culture was recommended. If a prior culture showed resistance to NFN and CIP, the data mining process recommended NFN. We then compared the theoretical data mining directed therapy outcomes to the outcomes that would have been achieved using conventional empirical therapy. In the URO for patients with (NU Algorithm ) and without (NU Algorithm ) a prior culture the data mining derived algorithms theoretically outperformed conventional empirical therapy by 11.3% (fig. 2, table 2). The effect was observed in men and women without and with a prior culture during most of the study period. Analysis of data mining derived algorithms for IM and ED visits Figure 2. Predictive outcomes from hypothetical empirical treatment of UTI based on NU derived algorithm evaluated at Northwestern Urology department. Asterisk indicates p For women (A) and men (B) seen in URO we compared theoretical treatment outcomes with conventional therapy and data mining derived therapy strategies, NU Algorithm and NU Algorithm. Asterisk indicates data mining derived algorithms theoretically outperformed conventional empirical therapy. Dagger indicates for conventional empirical therapy in men we assigned CIP in all cases. indicates that based on the NU Algorithm, if prior culture is not present, antimicrobial with superior susceptibility from prior year, specific to site where urine culture was obtained, is prescribed empirically. For patients in URO this is NFN for all years in question.

5 DATA MINING FOR IMPROVED EFFICACY IN URINARY TRACT INFECTION THERAPY 2261 Table 2. Theoretical outcomes of UTIs treated with conventional and data mining derived strategies in the Urology department Prior Urine Culture Sample Size Theoretical Treatment Success Conventional NU Algorithm % Improvement* Male: No ( ) Yes ( ) Overall Female: No ( ) Yes ( ) Overall 1, Total: No ( ) 14, Yes ( ) Overall 2, * Statistically significant difference between conventional and NU algorithm derived therapy. was also performed (data not shown). We identified a statistically significant 4.9% improvement in women evaluated in the ED compared to those who received conventional therapy. However, there were no significant improvements for men in the ED or for any patients in IM. DISCUSSION Antimicrobial resistance varies by time, visit site, pathogen, host and historical resistance patterns. We found a significant difference in antimicrobial susceptibility depending on the department to which the patient initially presented. URO patients experienced more highly resistant organisms, perhaps due in part to different demographics. Here we demonstrate a proof of concept that the application of rules derived from data mining algorithms can achieve a significant improvement in overall efficacy in empirical antimicrobial therapy in specific patient populations. Using predictive modeling we showed an overall significant improvement of 11.3% in URO patients when using our proposed algorithm compared to the administration of conventional therapy. We believe that this improvement will increase as more data are mined over time, 11 resulting in further improved outcomes and decreased costs. These added benefits come with negligible additional costs and inconvenience to the physician or patient. Data mining derived algorithms are site specific not only at our institution, but also in each NU department. While our report focuses on results for data derived from URO visits, analysis of data derived for IM and ED was also performed (data not shown). Using identical methodology comparing conventional therapy we found some measurable but not statistically significant advantage. We hypothesize this is because our treatment algorithm is most beneficial when the host environment is complex (as with URO patients with a history of prior UTIs, multiple organisms etc) and, consequently, conventional empirical therapy is less effective. However, data mining techniques are in fact more robust as the data set increases in size and scope. 11 We believe that outcomes derived from data mining at sites other than the URO will become statistically significant as bacterial resistance increases and more data are accrued, leading to global and incremental improvements in efficacy and cost savings over time. Ineffective empirical therapy leads to increased morbidity because of nonresolution of symptoms and it increases the costs of UTIs. The 2 most commonly used drugs for uncomplicated UTI, TMP-SMX and the fluoroquinolones, have lost 10% to 30% of their efficacy in the last decade, and our data appear to support these numbers. Antimicrobial resistance complicates the treatment of UTIs, and is associated with greater patient morbidity, higher costs of reevaluation and re-treatment, higher rates of hospitalization and greater use of broader spectrum antimicrobials. 17 A cost based model developed by Le and Miller demonstrated that fluoroquinolones become less expensive when TMP-SMX resistance exceeds 22% in the community. 18 Although models such as these provide important information regarding the selection of empirical treatment, they do not consider other factors that may influence these decisions such as societal costs (eg days lost from work or school due to UTI) and the effect that antimicrobial use (particularly fluoroquinolones) has on antimicrobial susceptibility in the community. 2 Real-time data regarding pathogen resistance and antimicrobial efficacy in outpatient UTIs are anecdotal at best, and not linked to patient scenarios, clinical or microbiological therapy outcomes. Data on outpatient antimicrobial resistance patterns are scant and usually based on hospital antibiograms which have significant limitations. 2 A hospital antibiogram quantifies drug resistance within a specific location but does not link data to individual patients. Furthermore, antimicrobial resistance is more prevalent in the inpatient than in the outpatient setting. 19,20 Therefore, we tend to overestimate antimicrobial resistance in an ambulatory setting by at least 10% to 30%. 2,19,20 Data mining, on the other hand, continually seeks unexpected patterns among demographic parameters linked to individual patients, capturing information dynamically as data accumulate from one point to another. Our study demonstrates the theoretical advantages of data mining derived algorithms to improve empirical therapy of outpatient UTIs. Our study has several important limitations. The analysis is limited by the nature of predictive modeling. Our outcomes are hypothetical in nature and assume that a patient not

6 2262 DATA MINING FOR IMPROVED EFFICACY IN URINARY TRACT INFECTION THERAPY only receives the appropriate culture specific antimicrobial, but also that symptoms resolve after treatment with the appropriate drug. Our data are also limited by the fact that they are specific to a single site. Nevertheless, we believe that our method has the potential for widespread use nationally and in local communities with improved outcomes at a minimal cost, and can be tailored to a multitude of clinical questions. CONCLUSIONS Antimicrobial resistance patterns in outpatient UTIs are site specific, and dependent on time and various host factors. Data mining directed therapy significantly improved theoretically determined outcomes when compared to contemporary conventional therapies for URO outpatients and for female patients in the ED. This approach could improve the efficacy and cost-effectiveness of empirical antimicrobial therapy for outpatient UTIs at a minimal cost. Our current data mining derived algorithms are site specific and offer the most significant advantages in the urology patient population where the environment is complex and resistance is higher. Future work should focus on prospective evaluation of these hypotheses as well as expansion of the generalizability of data mining techniques to other patient populations. REFERENCES 1. Foxman B: Epidemiology of urinary tract infections: incidence, morbidity, and economic costs. Dis Mon 2003; 49: Miller LG and Tang AW: Treatment of uncomplicated urinary tract infections in an era of increasing antimicrobial resistance. Mayo Clin Proc 2004; 79: Schaeffer AJ: New concepts in the pathogenesis of urinary tract infections. Urol Clin North Am 2002; 29: Foxman B, Barlow R, D Arcy H et al: Urinary tract infection: self-reported incidence and associated costs. Ann Epidemiol 2000; 10: Foxman B: Epidemiology of urinary tract infections: incidence, morbidity, and economic costs. Am J Med 2002; 113: 5S. 6. Wallach J: Interpretation of Diagnostic Tests, 7th ed. Philadelphia: Williams & Wilkins Griebling TL: Urologic diseases in America project: trends in resource use for urinary tract infections in women. J Urol 2005; 173: Griebling TL: Urologic diseases in America project: trends in resource use for urinary tract infections in men. J Urol 2005; 173: Warren JW, Abrutyn E, Hebel JR et al: Guidelines for antimicrobial treatment of uncomplicated acute bacterial cystitis and acute pyelonephritis in women. Clin Infect Dis 1999; 29: Wagenlehner FM, Weidner W and Naber KG: An update on uncomplicated urinary tract infections in women. Curr Opin Urol 2009; 19: Han J and Kamber M: Data Mining: Concepts and Techniques, 2nd ed. Burlington, Massachusetts: Morgan Kaufmann Publishers Gupta K, Scholes D and Stamm WE: Increasing prevalence of antimicrobial resistance among uropathogens causing acute uncomplicated cystitis in women. JAMA 1999; 281: Brown PD, Freeman A and Foxman B: Prevalence and predictors of trimethoprim-sulfamethoxazole resistance among uropathogenic Escherichia coli isolates in Michigan. Clin Infect Dis 2002; 34: Kahlmeter G: An international survey of the antimicrobial susceptibility of pathogens from uncomplicated urinary tract infections: the ECO.SENS Project. J Antimicrob Chemother 2003; 51: Cizman M, Orazem A, Krizan-Hergouth V et al: Correlation between increased consumption of fluoroquinolones in outpatients and resistance of Escherichia coli from urinary tract infections. J Antimicrob Chemother 2001; 47: Daza R, Gutierrez J and Piedrola G: Antibiotic susceptibility of bacterial strains isolated from patients with community-acquired urinary tract infections. Int J Antimicrob Agents 2001; 18: Hooton TM, Besser R, Foxman B et al: Acute uncomplicated cystitis in an era of increasing antibiotic resistance: a proposed approach to empirical therapy. Clin Infect Dis 2004; 39: Le TP and Miller LG: Empirical therapy for uncomplicated urinary tract infections in an era of increasing antimicrobial resistance: a decision and cost analysis. Clin Infect Dis 2001; 33: Richards DA, Toop LJ, Chambers ST et al: Antibiotic resistance in uncomplicated urinary tract infection: problems with interpreting cumulative resistance rates from local community laboratories. N Z Med J 2002; 115: Ti TY, Kumarasinghe G, Taylor MB et al: What is true community-acquired urinary tract infection? Comparison of pathogens identified in urine from routine outpatient specimens and from community clinics in a prospective study. Eur J Clin Microbiol Infect Dis 2003; 22: 242.