Electronic Algorithmic Prediction of Central Vascular Catheter Use
Objective. To develop prediction algorithms for the presence of a central vascular catheter in hospitalized patients with use of data present in an electronic health record. Such algorithms could be used for measurement of device utilization rates and for clinical decision support rules.
Design. Criterion standard.
Setting. John H. Stroger, Jr, Hospital of Cook County, a 464‐bed public hospital in Chicago, Illinois.
Participants. Patients admitted to the medical intensive care unit from May 31, 2005 through June 26, 2006 (derivation data set, May 31, 2005–September 28, 2005; validation data set, September 29, 2005–June 28, 2006).
Methods. Covariates were collected from the electronic medical record for each patient; the outcome variable was presence of a central vascular device. Multivariate models were developed using the derivation set and the generalized estimating equation. Three models, each with increasing database requirements, were validated using the validation set. Device utilization ratios and performance characteristics were calculated.
Results. Although Charlson score and duration of intensive care unit stay were significant predictors in all models, factors that indicated use or presence of a central line were also important. Device utilization rates derived from the algorithmic models were as accurate as those obtained using manual sampling.
Conclusions. Automated calculation of central vascular catheter use is both feasible and accurate, providing estimates statistically similar to those obtained using manual surveillance. Prediction modeling of central vascular catheter use may enable automated surveillance of bloodstream infections and enhance important prevention interventions, such as timely removal of unnecessary central lines.
Received March 4, 2009; accepted July 22, 2009; electronically published November 16, 2009.
Central line–associated bloodstream infections (BSIs) are common adverse events in healthcare centers1‐3 and result in substantial cost and patient morbidity. As part of a national emphasis on improving healthcare quality, rates of healthcare‐associated infection (HAI) have been proposed as quality measures for interhospital comparisons.4,5 Central line–associated BSI rates are a good measure of a hospital’s infection control practices, because these infections may be preventable. For example, most central line–associated BSIs could be prevented by adherence to recommended procedures during line insertion, removal of unnecessary lines, adequate preparation of catheter hubs, and maintenance of good insertion‐site dressings.6‐9
Automating HAI detection through use of electronic databases may facilitate public reporting of HAI rates. Although automated detection of HAIs is feasible, likely improves sensitivity, and reduces time and resources needed for surveillance,10‐13 manual collection of a denominator (ie, device‐days) is required to calculate a standardized rate. For central line–associated BSIs, central line–days are used to standardize rates and to risk‐adjust for comparisons between hospitals.2,14 Alternative strategies to daily manual measurement of central line use include estimation through random sampling,15 longitudinal systematic sampling with weekly or monthly enumeration,16 and use of patient‐days instead of central line–days.17 Automated electronic detection using procedure codes18 or natural language processing of radiological reports19 has also been accomplished with varying success.
We examined the accuracy of a statistical model to estimate the number of central line–days with use of electronic clinical data typically available in a clinical data repository. Such a model has 2 potential uses: (1) estimation of the denominator for central line–associated BSI surveillance and (2) detection of central line use at the patient level, which could facilitate a major central line–associated BSI–prevention effort by identifying candidates for central line removal.6
Methods
Setting and Participants
The present study was conducted at John H. Stroger, Jr, Hospital of Cook County, a 464‐bed public hospital in Chicago, Illinois. This hospital is the main safety‐net hospital of Cook County, serving as the central referral medical center for surrounding county clinics and hospitals. The study population was restricted to patients admitted to the 22‐bed medical intensive care unit (MICU) from May 31, 2005, through June 26, 2006; this cohort was derived from a study of the effectiveness of chlorhexidine bathing in preventing central line–associated BSI.20 All patients were assessed while in the MICU; patients were not followed up after transfer to the general medical floor. The present study was reviewed by the Institutional Review Board, and the need for informed consent was waived.
Data Collection and Variable Selection
For all patients admitted to the MICU, central line use was recorded prospectively each day. Each patient on each day was coded as either having a central line (eg, triple lumen catheter, peripherally inserted central catheter [PICC], or tunneled hemodialysis catheter) or not having a central line. For all patients, data resulting from clinical care automatically populated a clinical data repository.21 The data repository contained demographic information (ie, age, sex, race, and duration of hospital stay and MICU stay [the number of days from admission to the hospital or the MICU, respectively, to record sampling for presence of a central line]); comorbidities, as expressed by the Charlson score (calculated from discharge International Classification of Diseases, Ninth Revision (ICD‐9), admission codes obtained during admissions previous to a current admission22); medication orders; and laboratory and radiology use (ie, numbers and types of tests ordered). The clinical data repository is systematically reviewed for accuracy and completeness; missing data, when found, are replaced as needed. Therefore, data sets for the relevant period were considered to be complete. Vital signs and physician notes are not captured in our clinical data repository and were therefore unavailable for our predictive models.
We selected potential predictors for the presence of a central line with use of the following methods: clinician opinion about electronically recorded events frequently associated with the presence of a central line, medical record review of a sample of MICU patient records (
), and predictors identified using a data mining approach.23 From the manual medical record review, we identified the following medications as the most frequently used in MICU patients with a central line: total parenteral nutrition, lepirudin, oxybutynin, calcitriol, adenosine, tamsulosin, oxymetazoline, minoxidil, ketamine, bivalirudin, mycophenolate, fentanyl, ferrous sulfate, esomeprazole, and allopurinol. From the data mining approach, the following predictors were found: time from MICU admission to central line assessment sampling, number and type of intravenous medications (eg, vasopressors and antimicrobials), number of blood cultures and nonmicrobiology blood tests, number of chest radiographs (a potential indicator of central line placement), and use of specific radiological studies believed to require the presence of a central line (ie, angiograms, venograms, and other interventional radiological studies). Finally, clinician opinion resulted in the inclusion of Charlson score to assess comorbidities and demographic variables. Data for predictor variables were limited to those recorded within the 2 days before the day of assessment and were aggregated by patient into counts and mean values for inclusion in the logistic regression model. The data for these variables were collected from observations that occurred during the 48‐hour period before the observation of the presence or absence of a central line.
Development of Predictive Model and Statistical Analysis
From our study cohort, we created 2 data sets (a derivation and a validation data set). The derivation data set consisted of central line use observations that occurred in the MICU from May 31, 2005, through September 28, 2005. For the validation data set, we used every patient‐day during a 9‐month period (September 29, 2005, through June 28, 2006).
To find predictors of central line use, we performed bivariate and multivariable analyses by using the derivation data set. To account for the varying availability of data in clinical data warehouses, we developed 3 models that increased in the complexity of data requirements, as follows: a limited model, based on ICD‐9 codes, demographic characteristics, and duration of MICU stay (time from MICU admission to central line observation); a typical model, defined as the limited model plus laboratory and microbiological data; and a full model, defined as the typical model plus pharmacy and radiological data. Variables were recoded into categories when inspection of the data suggested deviation from normal distributions. Variables with 
were included in fully saturated multivariable models; variables were eliminated using a backward elimination process. We constructed parsimonious models to improve the generalizability of our work by simplifying data preparation. The generalized estimating equation with a binary outcome variable and an autoregressive correlation structure was used for significance testing in bivariate and multivariable analyses. Discrimination of the 3 models was assessed using the c statistic; calibration was assessed qualitatively using a plot of comparison of expected probabilities with actual probabilities of having a central line for each decile of expected outcome.
Validation of the Predictive Model
The entire cohort from the 9‐month data set was used as a validation data set. Every patient‐day in the MICU was considered to be eligible, and the 3 predictive models were applied to all available patient‐days in the validation data set. We used all available patient‐days for 2 reasons: (1) to assess the performance of the models in a manner that might be used prospectively and (2) to assess the performance of the models by using all available electronic information. The sum of products of covariates and β‐coefficients and the model intercept yielded the logit for each record in the data set. The exponentiation of the logit divided by the sum of one plus the exponentiation of the logit was calculated for each patient‐day to determine the predicted probability of central line presence for each patient‐day. The following 2 comparisons were made: (1) a population‐level assessment of aggregate device use as a proportion of total patient‐days, expressed as the device utilization ratio, and (2) performance characteristics of the 3 predictive models, expressed as the sensitivity and specificity. We also calculated the c statistic for each model in the validation set. For estimation of the population‐level central line utilization ratio, we summed the predicted probabilities of central line presence and divided by the total patient‐days for the cohort.
We calculated the device utilization ratio for a hypothetical weekly or monthly count by randomly sampling 1 day for each of the prespecified periods; the sampled day could be any day of the week or month. We calculated the point estimate and 95% exact binomial confidence intervals (95% CIs).
We calculated the differences between the known device utilization ratio (ie, prospective manual bedside review for each patient‐day), and we used the following methods for estimating the device utilization ratio: the summed probabilities from our validation data set (ie, a fully automated method) and weekly or monthly random samples of the previously recorded manual determinations. Analyses were conducted using SAS, version 9.1 (SAS Institute).
Results
Of 1386 eligible patient‐days among 257 unique patients in the MICU from May 31, 2005, through September 28, 2005, 854 occurred during the first week of the patient’s MICU stay, 296 occurred during the second week, and 236 occurred after the second week. A total of 1042 central lines were used in the cohort (mean, 1.2 central lines per patient among those with a central line); 503 (48%) were placed in femoral sites, 310 (30%) were placed in the internal jugular, 223 (21%) were placed in subclavian sites, and 6 (1%) were PICCs placed in the antecubital fossa. Temporary hemodialysis catheters represented 69 (7%) of the catheters, and the remaining catheters were triple‐lumen catheters or PICCs.
Predictors of central line use in the derivation set that were found during bivariate and multivariable analyses are listed in the Table. Although Charlson score and duration of MICU stay were significant predictors in all models, use of intravenous access through either number of blood samples obtained and processed, number of medications administered, or number of invasive radiological tests (and the frequent need for use of radiological contrast or access to vasculature) were important predictors in models that used additional data sets. Models were bounded at the lower range of predicted probabilities and overestimated line use at these ranges; at higher rates of device use, models underestimated use (Figure 1).
Figure 1. Plot of the calibration of predictive models for central line use. Points are generated by summing the probabilities in each decile of rank of prediction for each model and by plotting the actual prevalence of line use against this value.
The validation data set contained 3078 patient‐days of observation among 573 patients. One or more central line was present in 1889 patient‐days, yielding a device utilization ratio of 0.61. Among patients who had a central line, the mean number of lines was 1.2. A total of 2220 central lines were observed over the 3078 patient‐days; 823 (37%) were placed in femoral sites, 808 (36%) were placed in the internal jugular, 542 (24%) were placed in subclavian sites, and 27 (1%) were PICCs placed in the antecubital fossa. One hundred thirty‐one catheters (6%) were temporary hemodialysis catheters, and the remaining catheters were either triple‐lumen catheters or PICCs.
The point estimates for the number of central line–days were as follows: prospective bedside manual enumeration (catheter‐days, 1876; eligible patient‐days, 3054), weekly manual sample (catheter‐days, 269; patient‐days, 440), monthly manual sample (catheter‐days, 76; patient‐days, 108), electronic full model (catheter‐days, 1849; patient‐days, 3054), electronic typical model (catheter‐days, 1818; patient‐days, 3054), and electronic limited model (catheter days, 1704; patient‐days, 3054). Figure 2 shows the accuracy of the 3 electronic models and the 2 manual sampling strategies (weekly or monthly), compared with manual daily calculation of the device utilization ratio. The electronic full and typical predictive models and weekly sampling accurately estimated the device utilization ratio; the electronic limited model underestimated the ratio, and the monthly sampling strategy overestimated the ratio. All electronic methods produced smaller 95% CIs around the device utilization ratio estimate.
Figure 2. Population‐level central line utilization ratios among medical intensive care unit patients. Comparison of prospective daily manual bedside determination (data point and vertical line) with ratios estimated by predictive models by using either electronic data or manual sampling methods. Weekly count and monthly count are weekly and monthly sampling strategies, respectively. The x axis shows the rate difference between estimation methods and the gold standard daily count method. The device utilization ratio (DUR) is the ratio of central vascular catheter–days to total patient‐days in a unit.
Figure 3 shows the performance characteristics of the 3 predictive rules at the patient level for a variety of probability cutoff values. Use of a cutoff value of 0.25 maximized sensitivity at a cost of reduced positive predictive value. Both the typical and full predictive models at a cutoff value of 0.5 had both a sensitivity and positive predictive value greater than 70%; use of administrative data and day of MICU stay alone (ie, the limited model) dramatically reduced sensitivity when a probability cutoff value of less than 0.75 was used. The c statistics for the 3 models, as applied to the validation set, were 0.69 for the limited model, 0.74 for the typical model, and 0.76 for the full model.
Figure 3. Sensitivity and positive predictive value of automated central line detection algorithms for prediction in individual patients. Cutoff values indicate the probability threshold used to determine whether a line was present. The gold standard for performance characteristics was observed presence of a central vascular catheter on a patient‐day.
Discussion
We derived and validated 3 electronic models of increasing complexity for prediction of central line use in an MICU. Estimates of central line–days and device utilization ratios produced by these models performed well, generating device utilization estimates as accurate as those obtained through manual sampling. Weekly manual sampling yielded an accurate point estimate for the number of central line–days; however, an estimate generated using monthly sampling was less precise than those obtained using our electronic models.
The simplest electronic model was restricted to ICD‐9 codes and patient‐specific bed use, which are data commonly found in administrative data sets. An augmented second model was created by adding data on laboratory use—specifically, number of blood tests performed during the 48 hours before the date of bedside observation for the presence of a central line. The third and most complex model also included our pharmacy and radiology databases. In this model, we evaluated medications frequently associated with an MICU stay, central line access or intravenous medication administration, chest radiographs, and interventional radiographical studies that required the presence of a central line. The 2 models that incorporated data beyond administrative data were highly accurate and precise, predicting device use with 95% CIs within 5% of the daily observed rate (Figure 2). Furthermore, denominators resulting from prediction models, when used to calculate the incidence of central line–associated BSI, yielded rates markedly similar to those yielded using manual approaches. During the validation period, manual surveillance for central line–associated BSI that was conducted by the institution detected 6 episodes, yielding a measured rate of 3.2 episodes per 1000 central catheter–days. The use of the sampling or modeling methods would have yielded the following rates of central line–associated BSI: weekly sampling, 3.2 episodes per 1000 catheter‐days (95% CI, 3–3.5 episodes per 1000 catheter‐days); monthly sampling, 2.8 episodes per 1000 catheter‐days (95% CI, 2.5–3.2 episodes per 1000 catheter‐days); full model, 3.6 episodes per 1000 catheter‐days (95% CI, 3.4–3.7 episodes per 1000 catheter‐days); typical model, 3.4 episodes per 1000 catheter‐days (95% CI, 3.3–3.6 episodes per 1000 catheter‐days); and limited model, 3.6 episodes per 1000 catheter‐days (95% CI, 3.5–3.8 episodes per 1000 catheter‐days). The close similarity in rates suggests the potential applicability of this approach.
These models provide a framework for an evaluation of the use of electronic data to determine overall central line use in units and hospitals. Although calculation of comorbidities, as used in the limited model, provides some predictive capacity when combined with duration of hospitalization information, prediction of central line use appears to be enhanced when billing and financial data are augmented with more‐granular electronic databases, such as laboratory information systems and pharmacy and radiological databases.
The National Quality Forum includes central line–associated BSI among other HAI measures that are intended to indicate healthcare quality in hospitals.24 Simplifying surveillance may facilitate uptake of increasing requirements to monitor HAIs for quality improvement. Also, improvement of the efficiency of central line–associated BSI surveillance could redirect resources from manual surveillance to intervention efforts. The Centers for Disease Control and Prevention are committed to simplified methods for nosocomial infection surveillance for the National Healthcare Safety Network.25 Using data reported to this network, Klevens et al proposed the use of a weekly or monthly sampling strategy for manual surveillance of central lines as an acceptable alternative to daily monitoring.16 Our data also indicate that sampling would have been an acceptable alternative to manual daily enumeration.
In addition to aiding automated surveillance efforts, electronic models for vascular catheter detection provide theoretical benefits. Although insertion practice and line‐dressing placement are important elements in prevention of central line–associated BSI, excessive central line use and a failure to remove unnecessary lines are also modifiable causes of central line–associated BSI (Figure 4). Use of an electronic model to detect central lines provides the opportunity for intervention through the use of decision support rules. Rules could be developed that would identify patients who had unnecessary lines and could trigger alerts for central line infection control and removal efforts. For example, rules that detect central lines with a high positive predictive value could be used to generate a list of patients as candidates for infection‐prevention interventions, such as review of dressing quality. Alternatively, as electronic medical records evolve to capture central line use in individual patients, prediction models could be applied to patients in whom lines are noted to be present in the electronic medical record, and discordant results between the prediction rule and the observed line status might identify patients who no longer need a central line (eg, patients transferred from the MICU to the general medical floor or those with substantial improvements in their condition). To identify unnecessary central lines, a highly sensitive rule would be most useful, because patients without electronic prediction of central line use but in whom a vascular device is present would be targets of an intervention.
Figure 4. Conceptual model for risk adjustment of rates of central line–associated bloodstream infection (CLABSI). Total central line use does not account for unnecessary line use as an important modifiable suboptimal infection control practice.
Models for electronic prediction of central lines could not only eliminate the labor involved in collecting central line–days but could also provide more‐appropriate risk adjustment. At present, central line–days are used to risk‐adjust rates of central line–associated BSI, with the assumption that a central line–day represents a patient’s intrinsic risk of a primary BSI. A flaw of using central line‐days to risk‐adjust rates is that a component of total vascular device use is inappropriate central line use or vascular device use in individuals who no longer require a device (Figure 4). The current practice of risk adjustment using total device use adjusts for both appropriate and inappropriate device use. We posit that prediction models could be developed that are more likely to predict appropriate central line use. To gain a more accurate measure of healthcare quality, central line–associated BSI rates weighted for appropriate line use would provide a clearer understanding of the quality of an institution’s infection control practices. Although appropriateness could be measured using daily manual review, the resource requirements needed for such effort limit feasibility. Electronic prediction of appropriateness based on medication use and comorbidities, on the other hand, is achievable through the use of information technology.
The development of our central line prediction models was subject to several limitations. First, the data represent information from 1 MICU at 1 healthcare center, and the generalizability of results to other healthcare centers needs to be tested. In particular, many differences may exist between units and healthcare centers, such as characteristics of patients with central lines, clinical practices (eg, phlebotomy and microbiological culture), and prescribing practices (as assessed in the full model). It is likely that repeat validation of the model in multiple ward types, including general medical floors and intensive care units other than MICUs, and in multiple types of healthcare centers, including long‐term care facilities and acute care hospitals of varying specialties, would aid in creation of a robust model that permits accurate denominator prediction.
Second, it is likely that more‐complex models may be possible as more data become available electronically or are rendered computable through natural language processing (eg, radiological reports). Finally, we propose several potential uses of our models, and the variables used for model building should vary depending on the intended application. For example, Charlson score could be used to estimate total central line–days and to better adjust population‐level infection risk; however, for patients without records of medical history at a medical center, Charlson scores are not available until after ICD‐9 codes are generated after discharge, for future admissions. Until data sharing between centers through electronic health records or electronic documentation of diagnoses at admission is generally available, a Charlson score would not be calculable at admission for some individuals and could not be used in rules to determine whether placement of a central line is appropriate.
In conclusion, automated calculation of a central line utilization ratio with use of electronic data is both feasible and accurate, providing estimates statistically similar to those obtained using daily manual surveillance and likely preferable to weekly and monthly manual sampling. In theory, these calculations may provide a better risk adjustment for interhospital comparisons of rates of central line–associated BSI and overall quality of healthcare. Furthermore, prediction modeling of central line use offers the opportunity to enhance important prevention interventions, such as timely removal of unnecessary central lines.
Acknowledgments
Financial support. Centers for Disease Control and Prevention (CDC) cooperative agreement (1 U01 CI 000327–1 [CDC Prevention Epicenter Grant]).
Potential conflicts of interest. The authors report no conflicts of interest relevant to this article.
References
- 1. Kohn LT, Corrigan JM, Donaldson MS. To err is human: building a safer health system. Committee on quality of health care in America, Institue of Medicine report. Washington, DC: National Academy of Press; 2000.
- 2. O’Grady NP, Alexander M, Dellinger EP, et al. Guidelines for the prevention of intravascular catheter‐related infections. Centers for Disease Control and Prevention. MMWR Recomm Rep 2002; 51:1–29.
- 3. Wenzel RP, Edmond MB. The impact of hospital‐acquired bloodstream infections. Emerg Infect Dis 2001; 7:174–177.
- 4. Gaynes RP, Solomon S. Improving hospital‐acquired infection rates: the CDC experience. Jt Comm J Qual Improv 1996; 22:457–467.
- 5. Gaynes RP, Platt R. Monitoring patient safety in health care: building the case for surrogate measures. Jt Comm J Qual Patient Saf 2006; 32:95–101.
- 6. Pronovost P, Needham D, Berenholtz S, et al. An intervention to decrease catheter‐related bloodstream infections in the ICU. N Engl J Med 2006; 355:2725–2732.
- 7. Reduction in central line‐associated bloodstream infections among patients in intensive care units—Pennsylvania, April 2001‐March 2005. MMWR Morb Mortal Wkly Rep 2005; 54:1013–1016.
- 8. Marschall J, Mermel LA, Classen D, et al. Strategies to prevent central line‐associated bloodstream infections in acute care hospitals. Infect Control Hosp Epidemiol 2008;29(Suppl 1):S22–S30.
- 9. Yokoe DS, Mermel LA, Anderson DJ, et al. A compendium of strategies to prevent healthcare‐associated infections in acute care hospitals. Infect Control Hosp Epidemiol 2008;29(Suppl 1):S12–S21.
- 10. Panackal AA, M’Ikanatha NM, Tsui FC, et al. Automatic electronic laboratory‐based reporting of notifiable infectious diseases at a large health system. Emerg Infect Dis 2002; 8:685–691.
- 11. Effler P, Ching‐Lee M, Bogard A, Ieong MC, Nekomoto T, Jernigan D. Statewide system of electronic notifiable disease reporting from clinical laboratories: comparing automated reporting with conventional methods. JAMA 1999; 282:1845–1850.
- 12. Wright MO, Perencevich EN, Novak C, Hebden JN, Standiford HC, Harris AD. Preliminary assessment of an automated surveillance system for infection control. Infect Control Hosp Epidemiol 2004; 25:325–332.
- 13. Trick WE, Zagorski BM, Tokars JI, et al. Computer algorithms to detect bloodstream infections. Emerg Infect Dis 2004; 10:1612–1620.
- 14. Nosocomial infection rates for interhospital comparison: limitations and possible solutions: a report from the National Nosocomial Infections Surveillance (NNIS) System. Infect Control Hosp Epidemiol 1991; 12:609–621.
- 15. Cookson ST, Ihrig M, O’Mara EM, Hartstein AI, Jarvis WR. Use of an estimation method to derive an appropriate denominator to calculate central venous catheter‐associated bloodstream infection rates. Infect Control Hosp Epidemiol 1998; 19:28–31.
- 16. Klevens RM, Tokars JI, Edwards J, Horan T. Sampling for collection of central line‐day denominators in surveillance of healthcare‐associated bloodstream infections. Infect Control Hosp Epidemiol 2006; 27:338–342.
- 17. Tokars JI, Klevens RM, Edwards JR, Horan TC. Measurement of the impact of risk adjustment for central line‐days on interpretation of central line‐associated bloodstream infection rates. Infect Control Hosp Epidemiol 2007; 28:1025–1029.
- 18. Wright SB, Huskins WC, Dokholyan RS, Goldmann DA, Platt R. Administrative databases provide inaccurate data for surveillance of long‐term central venous catheter‐associated infections. Infect Control Hosp Epidemiol 2003; 24:946–949.
- 19. Trick WE, Chapman WW, Wisniewski MF, Peterson BJ, Solomon SL, Weinstein RA. Electronic interpretation of chest radiograph reports to detect central venous catheters. Infect Control Hosp Epidemiol 2003; 24:950–954.
- 20. Bleasdale SC, Trick WE, Gonzalez IM, Lyles RD, Hayden MK, Weinstein RA. Effectiveness of chlorhexidine bathing to reduce catheter‐associated bloodstream infections in medical intensive care unit patients. Arch Intern Med 2007; 167:2073–2079.
- 21. Wisniewski MF, Kieszkowski P, Zagorski BM, Trick WE, Sommers M, Weinstein RA. Development of a clinical data warehouse for hospital infection control. J Am Med Inform Assoc 2003; 10:454–462.
- 22. Deyo RA, Cherkin DC, Ciol MA. Adapting a clinical comorbidity index for use with ICD‐9‐CM administrative databases. J Clin Epidemiol 1992; 45:613–619.
- 23. Zogorski B, Trick WE, Solomon S, Cordell R, Weinstein RA. Comparison of predictive models to detect central venous catheters using data stored in the hospital information system. In: Program and abstracts of the 12th Annual Meeting of the Society for Healthcare Epidemiology of America (Salt Lake City, UT). 2002.
- 24. National Quality Forum (NQF). MQF endorsed standards: NQF #0139. Available at: http://www.qualityforum.org/measures_list.aspx. Accessed November 6, 2009.
- 25. Tokars JI, Richards C, Andrus M, et al. The changing face of surveillance for health care‐associated infections. Clin Infect Dis 2004; 39:1347–1352.
-
Presented in part: 45th Annual Infectious Diseases Society of America Meeting, San Diego, CA, October 4–7, 2007 (abstract 799).