A systematic review of validated methods for identifying transfusion-related sepsis using administrative and claims data


R. Carnahan, Department of Epidemiology, The University of Iowa College of Public Health, C 21-H General Hospital, 200 Hawkins Dr., Iowa City, IA 52242. E-mail: ryan-carnahan@uiowa.edu



To systematically review algorithms to identify transfusion-related sepsis or septicemia in administrative data, with a focus on studies that have examined the validity of the algorithms.


A literature search was conducted using PubMed, the database of the Iowa Drug Information Service (IDIS/Web), and Embase. A Google Scholar search was conducted because of difficulty identifying relevant studies. Reviews were conducted by two investigators to identify studies using data sources from the USA or Canada, because these data sources were most likely to reflect the coding practices of Mini-Sentinel data sources.


No studies that were identified that used administrative data to identify sepsis or septicemia related to transfusion of blood products. Thus, four studies that studied the validity of algorithms to identify sepsis and two that studied algorithms to identify allogeneic blood transfusion are described in this review. Two studies found acceptable positive predictive values of 80% and 89% for algorithms to identify sepsis in hospitalized patients. One study reported a negative predictive value of 80% in hospitalized patients, and another, a sensitivity of 75%. One study of veterans receiving surgery reported much worse performance characteristics. Two studies reported near-perfect specificity of codes for allogeneic red blood cell transfusion, but sensitivity ranged from 21% to 83%.


There is no information to assess the validity of algorithms to identify transfusion-related sepsis or septicemia. Codes to identify sepsis performed well in most studies. Algorithms to identify transfusions need further research that includes a broader range of transfusion types. Copyright © 2012 John Wiley & Sons, Ltd.


The US Food and Drug Administration (FDA) commissioned systematic reviews to identify validation studies of algorithms to identify 20 health outcomes of interest (HOIs) in administrative and claims data (hereafter “administrative data”), as part of its Mini-Sentinel program. These reviews provide the foundation for future studies of HOIs in Mini-Sentinel and other administrative data sources. In such studies, it is extremely important to understand the performance characteristics of the codes that might be used to identify an HOI, as the presence of a code is not always sufficient to determine that an HOI actually occurred. Because of FDA interest in studying the safety of blood products, one HOI selected for a systematic review was transfusion-related sepsis or septicemia.

Sepsis or septicemia can occur after the transfusion of a contaminated blood product. Historically, platelets have been the most likely blood product to be contaminated with bacteria, because they are stored at room temperature. Though platelets for transfusion are routinely screened for bacterial contamination, there is still risk that contaminated products can be transfused.[1] The source of an infection is not always obvious in a clinical setting, because people receiving transfusions are often quite ill and have other risk factors for sepsis. This is likely to pose challenges in confirming transfusions as a source of sepsis in administrative data.

This manuscript provides an overview of the transfusion-related septicemia or sepsis algorithm review. The full report can be found at http://mini-sentinel.org/foundational_activities/related_projects/default.aspx.


Details of the methods for these systematic reviews can be found in the accompanying manuscript by Carnahan and Moores.[2] In brief, the base PubMed search was combined with the following terms to represent the HOI: (" sepsis" [Mesh] or " sepsis" [All Fields] or " septicemia" [All Fields]), and (" transfusion" [All Fields] or " Blood Transfusion" [Mesh]). A search of the citation database of the Iowa Drug Information Service (IDIS/Web) was also conducted. The details of the searches can be found in the full report on the Mini-Sentinel website. The PubMed search was conducted on June 23, 2010, and the IDIS/Web search on May 10, 2010. All searches were restricted to articles published in 1990 or after. Mini-Sentinel collaborators were also asked to help identify any relevant validation studies.

The abstract of each citation identified was reviewed by two investigators. When either investigator selected an article for full text review, the full-text was reviewed by both investigators. Agreement on whether to review the full-text or include the article in the evidence table was calculated using a Cohen's kappa statistic. If fewer than five studies were identified that performed validation of the algorithm, up to 10 algorithms used without validation were reported. The data in the evidence table were extracted by one investigator and confirmed by a second.

Because of the limited number of citations identified for this HOI, and the lack of validation studies identified, several Google Scholar searches were also conducted using the following search strings: (i) sensitivity sepsis ICD; (ii) Sensitivity sepsis “international classification of disease”; (iii). “predictive value” sepsis ICD; and (iv) “predictive value” sepsis “international classification of disease.” Several searches to identify transfusion validation studies were also conducted. Because of the many results identified in a number of these searches, only the top matches were reviewed by a single investigator. This was an exploratory and less systematic process that also involved searching the references of related manuscripts until it seemed that most relevant studies were likely to have been identified.


Literature searches and reviews

Of the 52 abstracts reviewed, 17 were selected for full-text review; 27 were excluded because they did not study the HOI, 7 were excluded because they were not administrative database studies, and 1 was excluded because the data source was not from the USA or Canada. Cohen's kappa for agreement between reviewers on inclusion versus exclusion of abstracts was 0.40.

Of the 17 full-text articles reviewed, 1 was included in the final evidence tables (initially excluded for not studying HOI, but later included because it validated sepsis with no transfusion requirement); 5 were excluded because they did not study the HOI, and 11 were excluded because they did not use an administrative database. Reviewers identified 1 citation for review from full-text article references, which was included in the final report because it validated transfusion codes although not sepsis codes. Cohen's kappa for agreement between reviewers on inclusion versus exclusion of full-text articles reviewed could not be calculated by sas version 9.2 (SAS Institute Inc., Cary, NC, USA) because one reviewer rejected all articles on the initial screening. The one article on which they disagreed used a National Surgical Quality Improvement Program (NSQIP) database, which one reviewer initially thought was an administrative database. NSQIP data are collected and entered by research/quality improvement personnel, so this article was excluded because it did not use an administrative database. Mini-Sentinel collaborators provided no studies relevant to this outcome.

Because no studies were identified that used administrative data to identify sepsis related to transfusion, the inclusion criteria were expanded to studies that conducted validation of algorithms to identify sepsis or transfusion. Of the six studies included in the tables, one was identified from the initial search strategy,[3] one was identified through references of articles that underwent full-text review,[4] two studies from three sources were identified through the Google Scholar search process,[5-7] and one was identified serendipitously when searching for another article by the same author.[8] A final study was identified through references in a manuscript provided by FDA topic experts.[9]

Summary of algorithms

Four studies were reviewed that determined the performance characteristics of codes for sepsis, and two were identified that did so for transfusion (Tables 1 and 2).

Table 1. Sepsis algorithm validation studies
CitationStudy Population and Time PeriodDescription of outcome studiedAlgorithmValidation/Adjudication Procedure, Operational Definition, and Validation Statistics
Eaton et al. (2002)[5]Seventy-two patients admitted to a large university hospital during a 6-month period with a 038.x code were cases for the validation sample, and controls were those without a 038.x code.SepsisSepsis was identified by an ICD-9-CM code of 038.x (septicemia).Sepsis was confirmed through medical record review. A confirmation of sepsis required the presence of known or suspected infection plus three or more signs of systemic inflammation. Severe sepsis also required at least one dysfunctional organ system.[10, 11]
Martin et al. (2003)[6]Although 038.x was the focus of the validation study, a larger set of codes was used to identify sepsis in the Martin et al. study. Details can be found in the full report.Sepsis was confirmed in 64/72 patients with a discharge diagnosis code of 038.x, giving a PPV of 88.9% (95% CI 81.6–96.2%).
The negative predictive value (NPV) of code 038.x was 80.0% (95%CI 67.8–93.2%).
The PPV of a discharge diagnosis code 038.x in identifying severe sepsis only (sepsis + organ dysfunction) was 63% (95%CI 52–74%).
Another validation analysis defined sepsis as a systemic inflammatory response syndrome and acute organ dysfunction (the accepted clinical definition), without the requirement for infection. This increased the PPV of the 038.x code to 97.7% (95%CI 93.9–100.0%). The NPV remained 80.0%.
In another analysis, the PPV for codes 038.x for identifying patients with a systemic inflammatory response syndrome was 99% (95%CI 96–100%).
Ollendorf et al. (2002)[7]The study sample included 122 hospitalized patients from 10 medical centers participating in a clinical trial for severe sepsis of presumed infectious origin.Severe sepsis was present in all patients. Diagnosis codes were reviewed to determine their sensitivity in identifying these patients.The following ICD-9 CM codes on hospital bills (UB-92 forms) were considered sepsis: 003.1, 020.2, 022.3, 036.2, 038.x, 054.5, or 790.7. This list includes a number of additional codes for specific types of septicemia that are not captured by the commonly used 038.x codes. Definitions can be found in the full report.Severe sepsis was confirmed by the presence of 5 specified clinical criteria. These were determined prospectively because this was a randomized controlled trial.
Ninety-two of 122 hospital bills for these septic patients included a code for sepsis, for a sensitivity of 75.4%
Fifteen of the remaining 30 bills included codes for both major systemic infection and organ failure. Of the 15 without codes for systemic infection and organ failure, 4 had major infection codes only, 9 had organ failure codes only, and 2 had no codes that might indicate the presence of sepsis.
Romano et al. 2008[8]The study sample for sepsis included 12,011 patients in 110 Veterans Affairs (VA) hospitals in fiscal year 2001 with an operating room procedure for elective surgery plus hospitalization ≥ 4 days. Sepsis was present in 75 patients according to NSQIP data, the reference standard.Post-operative sepsisThe original patient safety indicator algorithm for sepsis used ICD-9-CM codes of 038.x in any discharge diagnosis field to identify sepsis.For systemic sepsis to be coded in the NSQIP data (the gold standard), the primary physician or medical record must have stated that the patient had systemic sepsis within 30 days after the operation.
An alternative algorithm tested in this study used the following ICD-9-CM codes in any secondary discharge diagnosis to identify sepsis: 038.x, 998.0, 998.1, 785.59, 785.50, 785.5, or 785.52.Using NSQIP data as the reference standard, the following performance characteristics were determined for the original patient safety indicator algorithm (ICD-9-CM codes 038.x):

Sensitivity: 32% (95%CI 23–43%)

PPV: 44% (95%CI 31–47%)

The following performance characteristics were determined for the alternative algorithm with additional ICD-9-CM codes:

Sensitivity: 37% (95%CI 27–49%)

PPV: 45% (95%CI 33–57%)

Specificity of all algorithms studied for various outcomes was reported to be >99.1%

Scanlon et al. (2008)[3]Pediatric hospital discharges from 2003 to 2005 from 76 hospitals. All surgery patients aged 0–17 years (excluding neonates) with a hospital stay >4 days, without sepsis or infection on admission and without a principal diagnosis of infection, were eligible. These criteria were met by 174,038 patients, and 4,367 patients met criteria for post-operative sepsis. Chart review was conducted for 279 cases.New onset sepsis after surgeryA secondary diagnosis code for sepsis during a post-surgical hospitalization as described.Medical record reviewers were physicians or nurses with clinical experience from participating hospitals. The methods stated that a number of outcome-specific questions were developed by pediatric experts, but did not state specific validation criteria.
Sepsis ICD-9 CM codes included: 038.x, 785.52, 785.59, 995.91, 995.92, and 998.0.Postoperative sepsis was confirmed in 223/279 cases in which chart review was conducted, giving a PPV of 79.93%
Infection and surgery codes are available from the AHRQ Pediatric Quality Indicators Technical Specifications Appendices.[12]
Table 2. Transfusion algorithm validation studies
CitationStudy population and time periodDescription of outcome studiedAlgorithmValidation/adjudication procedure, operational definition, and validation statistics
Segal et al. 2001[4]The study sample included 716 randomly selected hospital admissions at a large academic medical center, 358 with a billing code for red blood cell transfusion (ICD-9 procedure code 99.04) and 358 controls with no code for a blood component (ICD-9 procedure codes 99.0–99.09). Controls were stratified by DRG to get approximately the same number from each DRG. The two 1-month time periods utilized were August 1998 and March 1999.Blood transfusionBlood transfusion was identified by ICD-9 procedure code 99.04 (allogeneic red blood cell transfusion).The hospital's blood bank database was reviewed to determine whether there was a record of transfusion. For patients who were not included in the blood bank database, the electronic medical record was reviewed.
Controls did not have any blood component procedure code (99.0-99.09).Nine patients were identified by ICD-9 procedure code 99.04, but their records did not have a revenue code indicating they'd been billed for transfusion. To determine the impact of these missing data, these patients were categorized as true negative then false negative. Sensitivity and specificity were again calculated based on these classifications.
Of 358 admissions in which red blood cells were transfused, 61 were not billed. Thus, the sensitivity of billing codes was 83% (95% CI 79-87%).
When the 9 patients who were given procedure code 99.04 but had no revenue code were considered true negatives, the specificity was 100%.
When these patients were considered false-negatives, the specificity was 97.5% (95%CI 96–99%).
Patients who were not billed for transfusion were less likely to have commercial insurance.
Romano and Mark (1994)[9]This study examined 2,579 California hospital discharge abstracts from July to December 1987. Thirty hospitals were randomly selected for participation, stratified across different types of hospitals. Patients from the 10 most common diagnosis related groups (DRGs), or 9 related DRGs, were selected.Blood transfusionBlood transfusion was identified by ICD-9 procedure code 99.04 (allogeneic blood transfusions). The authors examined both the sensitivity and specificity of abstracts that allowed for either 3 procedure codes (i.e., truncated at 3) or 25 procedure codes.A total of 87/2,579 patients received transfusions, as determined by a re-abstraction of the hospital records in which specific comorbidities and procedures were purposefully identified.
The sensitivity of procedure codes from original discharge abstracts truncated at 3 fields was 21%. When this was expanded to allow 25 fields in the original discharge abstract, the sensitivity increased to 31%.
The specificity was 100% regardless of the number of fields considered.

Validation studies for sepsis

The performance characteristics of algorithms for sepsis varied by study. The International Classification of Disease, ninth edition, Clinical Modification (ICD-9-CM) codes 038.x were consistently used to identify sepsis. A number of additional codes were included that varied by study.

Eaton et al.[5] and Martin et al.[6] reported a validation study at Emory University Hospital of the ICD-9-CM 038.x codes on discharge records for identifying sepsis. In 72 cases identified by the code during a 6-month period, the positive predictive value (PPV) was determined to be 89% for sepsis as defined by suspected or confirmed infection plus a systemic inflammatory response syndrome (SIRS). The PPV of the 038.x codes in identifying SIRS was 99%. The PPV was 63% for identifying severe sepsis, defined as sepsis plus acute organ dysfunction. The negative predictive value (NPV) for codes 038.x among controls who were patients admitted immediately before or after the patient with a sepsis code was 80.0%. Finally, when the usual clinical definition of sepsis (SIRS plus acute organ dysfunction) was used, without requiring clear evidence of infection, the PPV of the 038.x code was 97.7%. Overall, the PPV of the 038.x code would be considered quite good, and NPV acceptable for most applications. The question becomes which patients with sepsis are less likely to receive a 038.x code, given that 20% is a substantial minority of the patients without a sepsis code who were septic. This study was limited in that it represents only a single institution's coding practices. It is also notable that an epidemiologic study reported concurrently by Martin et al.[6] included codes for septicemia (020.0), bacteremia (790.7), disseminated fungal infection (117.9), disseminated candida infection (112.5), and disseminated fungal endocarditis (112.81). It should be noted that the current ICD-9-CM code 020.0 represents “bubonic plague.” Code 020.2 represents “septicemic plague,” so it may be more appropriate for inclusion in an algorithm to identify septicemia. We are uncertain whether this discrepancy represents a change in code definitions or an error in the original manuscript. This study also provided a set of codes utilized to ascertain acute organ dysfunction, though no validation was conducted for any codes other than 038.x.

Ollendorf et al.[7] studied the sensitivity of a set of codes for identifying a group of 122 patients with severe sepsis of presumed infectious origin at 10 institutions. They were all participating in a clinical trial. In addition to various 038.x codes listed, this study utilized codes for anthrax septicemia (022.3), bacteremia not otherwise specified (NOS) (790.7), herpetic septicemia (054.5), meningococcemia (036.2), salmonella septicemia (003.1), and septicemic plague (020.2). The sensitivity of this set of codes for identifying sepsis was 75.4%. Of the 30 bills without a code for sepsis, 4 had major infection codes only, 9 had organ failure only, and 2 had no codes indicating sepsis. This study was unable to calculate PPV. The multicenter nature of the study may improve generalizability, but it is unclear whether the fact that these patients were in a clinical trial for severe sepsis may have changed the way their cases were documented and thus the selection of codes.

The two remaining studies of sepsis looked at surgical populations. Romano et al.[8] studied veterans by comparing administrative data to National Surgical Quality Improvement Program (NSQIP) data, which requires a chart diagnosis of sepsis to consider it present. This identified 75 patients with sepsis. The 038.x codes were first studied and found to have a sensitivity of 32% and PPV of 44%. An alternative algorithm that added six codes had a sensitivity of 37% and PPV of 45%. If considered among all NSQIP patients, the specificity was calculated as >99.1%. The alternative algorithm may be preferable in this patient population because it captured more cases without a decrease in positive predictive value, though the differences in algorithm performance were small. Overall the performance of these codes was less impressive than in the previous studies. This may be due to the post-surgical population, as they might have had other codes that took precedence. Alternatively, there may be different incentives to code sepsis properly in Department of Veterans Affairs (VA) facilities given their payment structure.

Scanlon et al.[3] studied a set of codes among surgical patients aged 0–17 years discharged from 76 pediatric hospitals, excluding neonates. Chart review was conducted for 279 patients with an ICD-9-CM code for sepsis. This study focused on new onset sepsis after surgery, as it was to assess pediatric hospital quality indicators. In addition to the 038.x ICD-9-CM codes, this study utilized codes for septic shock (785.52), shock without mention of trauma, other (785.59), postoperative shock (998.0), SIRS due to an infectious process without organ dysfunction (995.91), SIRS due to an infectious process with organ dysfunction (995.92). The PPV was 79.93%.

Overall, the PPV of the various sets of codes for sepsis was relatively high in every study except Veterans Affairs (VA) post-surgical patients.[3, 5-8] As stated, this difference may have something to do with the patient population or the payment structure of the VA that could influence coding practices. Two studies took place after the year 2000 and, thus, reflect relatively recent coding practices,[3, 8] whereas the specific dates of the other two studies were unclear.[5-7]

Validation studies for transfusion

Segal et al.[4] studied 358 patients with an ICD-9 procedure billing code 99.04 for allogeneic red blood cell transfusion and 358 controls without any billing code for a blood product, all from a single large academic medical center. The hospital's blood bank data was used to determine transfusion status, and electronic medical records were reviewed for patients without blood bank records. The sensitivity of the billing codes was 83%. Patients without commercial insurance were less likely to have a billing code for transfusion, perhaps reflecting an effect of reimbursement differences on the likelihood of receiving a code. A sensitivity analysis was conducted in which nine patients who received a procedure code 99.04 but not a revenue code were considered either true negatives or false negatives. When they were considered true negatives, the specificity of code 99.04 was 100%. When they were considered false negatives, the specificity was 97.5%. At least at this single academic medical center, code 99.04 appears to be relatively sensitive and highly specific. Commercial insurance may improve the sensitivity of billing codes.

In another study of 2,579 California hospital discharge abstracts from 1988, Romano and Mark[9] found that the sensitivity of ICD-9 procedure code 99.04 was only 21% when the top three procedure codes were reviewed (as were available in Medicare claims), and 31% when 25 procedure codes were reviewed. The specificity was 100%. This study reduces the confidence that procedure code 99.04 is sensitive for capturing transfusions, but confirms near perfect specificity. The study included 30 hospitals randomly selected across type of hospital (e.g., university teaching, non-teaching, small rural, etc.). Though the results are somewhat dated, they might be considered more generalizable than the single-center study by Segal et al.[4]


This review identified no studies using algorithms to identify transfusion-related sepsis or septicemia using administrative data. Such an algorithm and validation study might consider the temporal relationship between transfusion and sepsis, as well as the probability that sepsis might have developed because of other exposures such as surgery or trauma, both of which are common in patients who receive blood transfusions. In addition to the currently available codes, the FDA's Center for Biologics and Evaluation Research (CBER) has proposed new ICD-9-CM codes for identifying infections determined to be transmitted by blood transfusions.[13] A new code was adopted for acute infection following transfusion, infusion, or injection of blood and blood products (999.34), to be implemented on October 1, 2011.[14] Previously, blood product associated infections might have received an ICD-9-CM code 999.3 (complications of medical care, not elsewhere classified, other infection) or 999.39 (infection following other infusion, injection, transfusion, or vaccination). The new code 999.34 will add more specificity to the definition, though it does not state that the infection was transmitted by the blood or blood product, in contrast to the code originally proposed by CBER. This code would be used in combination with an additional code to describe the type of infection, such as 038.x to identify sepsis.

Several studies were identified that examined the validity of algorithms to identify sepsis or allogeneic red blood cell transfusion. Because of the variability in algorithms and study populations in the sepsis studies, it is difficult to make a recommendation for one algorithm over another. Codes 038.x appear to have acceptable performance characteristics to identify sepsis in most settings. There is no clear evidence to indicate whether the additional codes utilized would improve or worsen the balance of performance characteristics, though it would seem reasonable to consider the addition of other codes that represent septicemia. The study by Romano et al.[8] in veterans showed slightly better performance characteristics with the addition of extra codes, though no significant difference in performance was described. Given that this study found much worse algorithm performance characteristics than the other studies of sepsis, caution should be used in generalizing the results.

The code for allogeneic red blood cell transfusion was also found to have fairly good performance characteristics in a study conducted at one hospital,[4] but another study found relatively poor sensitivity.[9] The performance of other transfusion codes, including those for autologous blood donation or platelet transfusion, was not studied. One study of transfusion that used both ICD-9-CM procedure code 99.03 (transfusion of whole blood) and 99.04 (transfusion of allogeneic red blood cells), in addition to the “blood pints furnished” variable from MedPAR data sets, provides some useful information about the codes, despite a lack of validation studies on the “blood pints furnished” variable. Anderson et al.[15] examined blood use in elderly Medicare beneficiaries with an inpatient hospital stay during 2001. When either a procedure code or non-zero entry for blood pints furnished was used as the criteria for transfusion, only 77% of transfusion recipients had a procedure code for transfusion. A total of 36% of transfusion recipients had a nonzero blood pints furnished value, and only 13% had both a nonzero blood pints furnished variable and a procedure code for transfusion. Most blood centers charge for the transfusion preparation and procedure but not for the blood itself, and there is evidence of confusion in how to bill for blood transfusions that likely led to under-coding of transfusions in recent history. Thus, it is likely that transfusions are under-identified in administrative data, despite the fact that transfusions are highly likely to have taken place when transfusion codes are present in these data.

Platelets have historically had a higher rate of contamination with infectious organisms than other transfusion types because of room temperature storage, particularly prior to the practice of routine screening of platelets for contamination. Therefore, it would be useful to examine the performance characteristics of codes for platelet transfusion for surveillance of platelet-related infections.

Given the limited available information on the identification of sepsis specifically related to transfusion in administrative data, it is difficult to assess the potential to identify this outcome. Populations that receive transfusions are often at high risk of infection and sepsis because of other reasons (e.g., trauma or surgery). The most definitive diagnosis might take place if a specific organism is identified late in cultures of platelet samples, for example, taken prior to the transfusion, as all platelets are now screened for bacterial contamination.[1] Platelets may be contaminated with low levels of microorganisms from the outset, but growth during storage at room temperature may lead to higher levels when the transfusion is actually given. Various changes in blood collection and screening have decreased the risk, but none of these added measures have resulted in perfect safety.[1] Other infections might be deemed related to transfusion if no other potential source of the specific infectious organism is known, but true confirmation of the source of the infection can be difficult.

One strategy for improving specificity of algorithms might be to focus on organisms known to more commonly contaminate blood products, as has been described elsewhere.[16] However, there have also been some cases of emerging infectious diseases contracted from transfusion. The blood supply is not necessarily screened for these diseases, which have historically been confined to endemic areas outside the USA. Increasing incidence of infection with these diseases related to transfusion would support expansion of screening efforts to include them.[17]


Future research on sepsis code validation might focus on the performance of codes other than 038.x to identify sepsis or septicemia, such that an optimal combination of codes could be determined. If a specific algorithm is designed to identify sepsis that is caused by a transfusion, special attention will need to be paid to the most likely source of the infection insofar in that it can be determined. Patients who receive transfusions often have other risk factors for sepsis that need to be considered. It may be useful to study specific infectious organisms or other specific criteria that might implicate the transfusion in the development of sepsis. It might also be useful to explore the addition of ICD-9-CM code 999.3, 999.39, or the new code 999.34, to the algorithm to identify transfusion-related sepsis.

Transfusion codes other than 99.04 also have unknown performance characteristics, and even the performance of this code varied across studies. It would be helpful to examine the performance characteristics of the code for platelet transfusion in order to study infections related to this exposure. It would also be helpful to examine concordance of transfusion procedure codes and “blood pints furnished” revenue codes, and the relationship of each of these codes to actual transfusion, to determine whether algorithms should be expanded beyond procedure codes when revenue codes are available.


The authors report no conflicts of interest related to this work.


  • Algorithms to identify sepsis had positive predictive values of 80% or better in two out of three studies.
  • Two studies found essentially perfect specificity of a code to identify allogeneic red blood cell transfusion, but sensitivity was highly variable.
  • No studies were identified that validated codes for transfusion of other blood products, such as platelets.
  • It will likely be very difficult to confirm transfusions as a cause of sepsis using current administrative data. However, a code has been newly adopted for “acute infection following transfusion, infusion, or injection of blood and blood products,” which could add specificity to algorithms.


The authors would like to thank Mikhail Menis, Pharm.D., and colleagues at the US Food and Drug Administration's Center for Biologics Evaluation and Research for their input on this work. Mini-Sentinel is funded by the Food and Drug Administration (FDA) through Department of Health and Human Services (HHS) Contract Number HHSF223200910006I. The views expressed in this document do not necessarily reflect the official policies of the Department of Health and Human Services, nor does mention of trade names, commercial practices, or organizations imply endorsement by the US government.