A Survey of LOINC Code Selection Practices Among Participants of the College of American Pathologists Coagulation (CGL) and Cardiac Markers (CRT) Proficiency Testing Programs

The CAP LOINC study consisted of a questionnaire (see Data Supplement S1 of supplemental digital content, containing 1 table and 3 data supplements at www.archivesofpathology.org in the May 2020 table of contents) that was sent to US and Canadian participants in CAP's 2018 CGL and/or CRT Proficiency Testing surveys. The CGL survey encompasses 7 procedures: activated partial thromboplastin time (APTT), fibrinogen, international normalized ratio (INR), prothrombin time (PT), D-dimer (DD), fibrin(ogen) degradation products (FDP) in plasma, and fibrin(ogen) degradation products (FDP) in serum. The CRT survey encompasses 3 procedures: creatine kinase–muscle/brain (CK-MB) immunochemical, myoglobin, and troponin I (TnI). Table 1 describes the analytes, procedures, and specimen types for the CGL and CRT surveys.

Laboratories were asked to report details on the primary method and instrument used for testing each analyte, usage and application of LOINC codes, and the codes, if any, which are assigned to each procedure. The instrument and reagent for each analyte were selected from a predefined list provided by CAP as part of the standard external quality assessment procedure (see Data Supplement S2). The survey included questions asking laboratories whether the tests' assigned LOINC codes were used only in the laboratory information system (LIS) or if LOINC codes are sent from the LIS to other systems, including the EHR or the computerized physician order entry system (CPOE). In addition, respondents were asked to categorize the practice setting of the laboratory: nonacademic hospital/medical center, academic hospital/medical center (eg, has an accredited pathology residency program), government-owned facility (city, county, or state), Veterans Affairs (VA)/Department of Defense (DoD)/military treatment facility, physician office laboratory, independent reference laboratory (IRL), or other. All CGL and/or CRT participants received 1 reminder notice encouraging participation. The questionnaire results were anonymized with each respondent being assigned a unique identifier. No incentive was provided for completion of the questionnaire.

To analyze the attributes of the LOINC codes reported by the laboratories, 2 pathologists with expertise in LOINC and the laboratory domains of coagulation and cardiac testing independently reviewed all codes and their corresponding entries in the LOINC database, using LOINC version 2.65 (released December 2018; Regenstrief Institute Inc, Indianapolis, Indiana).¹²  All codes reported by the laboratories were assessed for version changes during the study period. The 6 LOINC axes/parts (component, property, time aspect, system, scale type, and method type), fields for the LOINC code status (active, deprecated, discouraged, trial), and order/observation classification (order, observation, or both) were reviewed for each LOINC code provided by respondents.

LOINC codes were evaluated for optimal selection for procedures where a specific LOINC code was more appropriate than other available codes. Additional information that a laboratory might use in selecting the appropriate LOINC code, based on the required sample type for a test, was obtained from the LOINC database's “related names 2” and “long common name” fields. The “related names 2” field is intended to facilitate users' finding their intended concept when searching the database and includes terms associated with the LOINC concept and related names. The “long common name” field is an algorithmically generated name that uniquely identifies that code in the LOINC database. The most appropriate LOINC code was defined as the code with the greatest specificity, accuracy, and detail for each procedure as designated by the 2 pathologists in addition to defining suboptimal (eg, less specific/accurate) codes; disagreements between both raters were resolved by a third pathologist. Each axis of the LOINC code was assessed independently for appropriateness, and codes with errors in any of the LOINC axes/parts or in the order/observation field were identified. Considerations for determining LOINC code accuracy, specificity, and errors included whether the LOINC code existed in the LOINC database; whether the component, sample type, and reporting unit(s) were consistent with the procedure reported in the CAP's CGL or CRT survey and the method/instrument reported by the laboratory; and whether the code represented the appropriate test (order/observation code for the appropriate individual analyte) versus a panel (order code that cannot be used to obtain an individual analyte).

Statistical analyses were performed using Stata version 15 (StataCorp, College Station, Texas). Graphical representations of the data were implemented in open source software (R version 3.4.2, Comprehensive R Archive Network). The Shannon entropy was used as a measure of dissimilarity or randomness for each of the 6 LOINC axes/parts for each test.¹³  Briefly, the relative frequencies of all categories, where S represents the number of categories within an individual LOINC axis/part, were determined for each test and represented as the proportions vector, p, with a sum of 1. The Shannon entropy (H) was then calculated as:

If one value of the vector p is 1 (ie, all observations belong to a single category), then H assumes a minimum value of 0 representing no dissimilarity or randomness. If all values of p are equal to 1/S (ie, each category has the same number of observations), then H assumes a maximum value of ln(S), representing the highest dissimilarity or randomness.¹⁴  Dissimilarity scores for each of the 6 LOINC axes/parts for each survey procedure were represented in a radar plot.

The survey was sent to 1916 participants of the 2018 CGL and/or CRT surveys. Of 1916 CGL and/or CRT participants, 90 (4.7%) responded to the questionnaire. Most respondents (n = 62, 68.9%) reported their laboratory setting. Responses from 2 nonacademic centers that indicated that respondents were unsure of which tests their laboratories performed and whether LOINC codes were used were excluded from analysis. Of the remaining 60 respondents, nonacademic hospital/medical centers comprised the largest proportion of respondents (n = 36, 60.0%), with 6 responses (10%) each coming from IRL, VA/DoD/military, and government-owned facilities; academic centers and office laboratory settings had the fewest respondents, with 5 respondents (8.3%) and 1 respondent (1.7%), respectively (see Figure 1).

Approximately one-third (n = 31, 34.4%) of the respondents indicated that they were unsure if 1 or more analytes had an assigned LOINC code in their LIS. Overall, 62 of 90 (68.9%) and 37 of 90 (41.1%) respondents self-reported that they do have an assigned LOINC code in the LIS for at least 1 analyte in the CGL and/or CRT surveys, respectively. All of the VA/DoD/military respondents indicated that they use LOINC codes in their LIS, and half of these laboratories also reported that LOINC codes are sent from the LIS to the EHR. Most nonacademic centers indicated that they use LOINC codes in their LIS (31 of 36, 86.1%) and that LOINC codes are sent from their LIS to the EHR (20 of 36, 55.6%). Five of the 36 nonacademic centers (13.9%) also reported using LOINC codes for CPOE. The IRL and government-owned laboratories had equivalent reported usage of LOINC codes in the LIS (5 of 6, 83.3%) and of LOINC codes sent from the LIS to the EHR (3 of 6, 50.0%); neither of these settings reported sending LOINC codes from the LIS to the CPOE. Most (4 of 5, 80.0%) academic center laboratories reported using LOINC codes in the LIS and most (3 of 5, 60.0%) also reported that LOINC codes are sent from the LIS to the EHR; only 1 academic laboratory (1 of 5, 20.0%) reported that LOINC codes are sent from the LIS to the CPOE. The distribution of self-reported LOINC usage within the LIS and the proportion of each practice setting that sends LOINC codes from their LIS to the EHR and CPOE are shown in Figure 1.

Of the 90 respondents, 47 (52.2%) provided data on the method, instrument, and/or reagent for at least 1 procedure in the CGL and/or CRT surveys. Table 2 summarizes the frequency of laboratories reporting 0 to 7 analytes in the CGL survey and 0 to 3 analytes in the CRT survey, as well as the percentage of respondents providing that number of analytes for each survey. The number of laboratories that reported the method and/or instrument for each analyte varied significantly, with the greatest number of laboratories providing the method and/or instrument for APTT and the fewest for FDP measured in serum: APTT (n = 43, 47.8%), fibrinogen (n = 36, 40.0%), INR (n = 35, 38.9%), PT (n = 38, 42.2%), D-dimer (n = 33, 36.7%), FDP in plasma (n = 10, 11.1%), FDP in serum (n = 6, 6.7%), CK-MB (n = 31, 34.4%), myoglobin (n = 16, 17.8%), and TnI (n = 31, 34.4%). In total, the method/instrument information was provided for 279 tests; of these, 266 (95.3%) had an associated LOINC code provided by the respondent. Of the total LOINC codes reported (n = 275), 9 (3.3%) did not have an associated method/instrument reported by the respondent; all but 2 codes (5934-2 and 12345-1) (0.7%) corresponded with codes that exist in the LOINC database. All codes reported by the laboratories were current with the codes in the LOINC V2.65 database and there were no version changes in the LOINC codes used during the study period. Of the laboratories that reported LOINC codes, all selected codes were active, and no deprecated or discouraged LOINC codes were supplied by any of the laboratories.

Although many laboratories reported using the same method or instrument/reagent, there was significant variability in the LOINC code selected by the different laboratories. The relationship between survey responses for method or instrument/reagent and LOINC code selection for each analyte in the CGL and CRT surveys is represented by individual chord diagrams in Figure 2, a through i. The arc length corresponds to the frequency of respondents in each group and the absolute number of respondents for each group is indicated by the numbers on the axis.

The most frequently used platform in the CGL survey (Instrumentation Laboratory, Bedford, Massachusetts) was used by 19 laboratories (21.1%) for 68 procedures and was listed as the instrument used to measure a maximum of 5 of the 7 CGL analytes (71.4%) by 8 laboratories (8.9%). The same LOINC code was reported by these laboratories for 14 of 15 APTT procedures (93.3%), 13 of 15 PT procedures (86.7%), 12 of 14 INR procedures (85.7%), and 9 of 11 fibrinogen procedures (81.8%); 4 of 10 D-dimer procedures (40.0%) had the same LOINC code reported, while another 4 of 10 (40.0%) had 1 of 2 other LOINC codes reported. There was less agreement among LOINC codes reported for CGL procedures for Stago (Diagnostica Stago, Asnières sur Seine, France) and Siemens (Siemens Healthineers, Erlangen, Germany) platform users (Figure 2).

The most frequently used platform in the CRT survey (Siemens) was used by 17 laboratories for 33 procedures and was listed as the instrument used to measure all 3 CRT analytes by 5 laboratories. The same LOINC code was reported by these laboratories for 70.0% (7 of 10) of CK-MB procedures, 100.0% (6 of 6) of myoglobin procedures, and 87.5% (14 of 16) of TnI procedures. Variability among LOINC codes reported for the 3 CRT procedures among other platform users is shown in Figure 2.

To characterize the nature of the discrepancy, the 6 LOINC axes/parts (component, property, time aspect, system, scale type, and method type) from the LOINC database were analyzed for the reported LOINC codes. The Shannon entropy was used to evaluate the degree of dissimilarity of the LOINC codes (defined by their 6 LOINC axes/parts) for each analyte in the CGL and CRT surveys. Figure 3 shows the degree of dissimilarity of each axis for each analyte. There was complete agreement for the system axis for fibrinogen. All other analytes had LOINC codes with variation in the system axis with the greatest dissimilarity indices for FDP in plasma and serum, APTT, PT, and INR. There were also differences in the component axis selected, particularly for CK-MB, D-dimer, and FDP in plasma, and the property axis selected, particularly for D-dimer, FDP in plasma and serum, and CK-MB (Figure 3). There was no dissimilarity in the time aspect and scale axes for all analytes except the APTT and PT (0.115 for APTT and 0.119 for PT, respectively, for both axes); dissimilarity for these 2 procedures were due to the selection of LOINC codes that did not exist in the LOINC database. Dissimilarity for the method axis was observed with the FDP in plasma, CK-MB, APTT, and PT procedures; again, the dissimilarity for the APTT and PT procedures were due to the selection of LOINC codes that did not exist in the LOINC database (Figure 3).

Descriptions of LOINC codes selected, and their frequencies, for each analyte in the CGL survey and the CRT survey are provided in Tables 3 and 4, respectively. For the 275 observations with LOINC codes reported, 54 (19.6%) were incorrect (Tables 3 and 4). The number of errors for each of the LOINC dimensions of component, property, time, system, scale, and method was 8.0% (n = 22), 9.8% (n = 27), 0.7% (n = 2), 9.8% (n = 27), 0.7% (n = 2), and 1.5% (n = 4), respectively (Tables 3 and 4); the most common errors in LOINC code selection involved the component, system, and property axes/parts. Suboptimal (eg, less specific/accurate) codes predominantly involved the system axis. All of the CGL procedures except the D-dimer and FDP in serum had at least 1 LOINC code reported with an incorrect component, while only the CK-MB procedure in the CRT survey had LOINC codes reported with an incorrect component (Table 4). For example, in the test code selection for APTT, 1 laboratory selected a LOINC code for activated clotting time (Table 3); for fibrin degradation products in plasma, 1 laboratory selected a LOINC code for plasmin inhibitor actual/normal (Table 3); and for CK-MB, 1 laboratory selected a LOINC code for TnI (Table 4). In addition to errors in component selection, other laboratories selected LOINC codes for panels. For example, in the LOINC code selection for INR, 1 laboratory selected PT panel (Table 3), and in the LOINC code selection for CK-MB, 1 laboratory selected creatine kinase panel (Table 4). LOINC codes were also selected with the incorrect property axis/part, which represents the unit type (eg, mass concentration, catalytic fraction). For example, although the CAP survey results form for CK-MB has 2 unit options, ng/mL (mass concentration) and U/mL (arbitrary concentration), laboratories selected LOINC codes with an incorrect unit property including catalytic fraction, catalytic concentration, and ratio (Table 4). Similarly, for INR, some laboratories incorrectly selected codes that had both the wrong component (coagulation tissue factor induced) and the wrong property (time) (Table 3).

To gain insight into what is inferred with the use of one property such as Bld (blood) as opposed to another property such as PPP (platelet-poor plasma), the “related names 2” and “long common name fields” were reviewed for each code. As demonstrated in Supplemental Table S1, these fields contain additional information that a laboratory might use in selecting the appropriate LOINC code based on the test and sample type. In reviewing the LOINC database “related names 2” field, probable mistakes within the database itself were identified. For example, in the “related names 2” field for D-dimer, several concepts such as “Rh,” “Rhesus,” and “Dextro” appeared to be erroneously associated with this component (Supplemental Table S1).

Errors in the sample type (system axis) varied in the degree of inaccuracy. For example, 1 laboratory selected control blood (Bld^control) as the sample type for APTT (Table 3). When selecting codes for PT, INR, and APTT, many laboratories selected a code for blood (Bld), whereas others selected a code for platelet-poor plasma (PPP). For the LOINC codes where the system was PPP, the additional fields contained the following words related to sample type: platelet-poor plasma and plas (abbreviation for plasma). For the LOINC codes where the system was blood (Bld), the additional fields contained the following words related to sample type: blood, whole blood, and WB (abbreviation for whole blood). This additional information conveys that Bld is not merely an umbrella term for blood-derived products, rather it is defined specifically as whole blood as indicated in section 2.5 of the LOINC User's Guide V2.65.¹⁵  Notably, blood is not included in the “related names 2” and “long common name fields” for the LOINC codes selected in the CGL that have PPP as the system. Furthermore, none of the LOINC sample type name options for blood in the LOINC database explicitly include the term whole (eg, whole blood) (Data Supplement S3).¹² 

Thirty-one laboratories (34.4%) provided a LOINC code for a quantitative, TnI cardiac assay performed on serum/plasma, and reported in mass concentration units. Of these, 5 (16.1%) reported using a LOINC code with the method column indicating that the detection limit is 0.01 ng/mL or lower (equivalent to 10.0 ng/L or 0.01 μg/L). The method or instrument/reagent selected by each of the 31 laboratories was reviewed. None of the 5 laboratories that selected the LOINC code with a method detection limit of 0.01 ng/mL or lower had reported using a high-sensitivity assay method in the CAP CRT survey.¹⁶  Four laboratories (12.9%) did report using a high-sensitivity troponin (HscTn) assay, but none of these laboratories chose the LOINC code with a method detection limit of 0.01 ng/mL or lower.

This study evaluated LOINC code selection for common laboratory tests across an array of laboratory settings. Review of respondents' data further revealed that many of the laboratories perform these tests with the same method or instrument/reagent. The use of common laboratory equipment to perform common tests raises the expectation that LOINC code selection would be uniform across laboratory environments. Such is not the case as this study demonstrates that even for these very common tests, there were significant differences between the LOINC codes selected by laboratories. Furthermore, there were numerous basic and fundamental errors demonstrated in the LOINC codes provided by respondents, including errors in LOINC code selection for all 4 analytes (APTT, PT, fibrinogen, and CK-MB) regulated for proficiency testing by CMS. Errors ranged from LOINC terms with the incorrect system (ie, specimen type) to LOINC terms representative of a completely different assay from the one actually performed. In fact, 1 or more laboratories chose a LOINC code with the incorrect component (ie, wrong test) for 6 of the 10 analytes in the survey (Table 3). Laboratories also selected LOINC codes for panels, which can never be mapped to an individual analyte and should have been automatically excluded when choosing the LOINC code. Such fundamental errors in LOINC code selection have the potential for multiple unintended consequences. Most concerning is the potential for laboratory data to be misinterpreted in a clinical setting as a result of an incorrectly encoded laboratory test. Less immediate issues pertaining to how LOINC-encoded laboratory data can, or cannot, be meaningfully aggregated, and management of varying levels of granularity, are equally problematic.¹⁷ 

For the analyzers used by the laboratories in this survey, coagulation tests such as APTT and INR are performed on PPP and are never performed on whole blood, so the selection of a code that indicates the system as PPP would more accurately represent the test performed as opposed to a code that lists the system as blood (Bld). Superficially, an end-user may think that “Bld” is a nonspecific answer for a collected blood specimen. However, LOINC has defined the system part term Bld as whole blood as indicated in section 2.5 of the LOINC User's Guide V2.65.¹⁵  Blood has the connotation or implied meaning of whole blood, as indicated in the “related names 2” and “long common name” data fields, which persons assigning a new LOINC code to a laboratory test would reasonably use to guide their code selection. In this instance, aggregation of general and granular concepts would not correct these errors. A method to view LOINC parts and terms as each relates to the other might have helped steer the coding staff to a more correct term selection. In its current form, LOINC does not have the logical infrastructure to enable logical inferences and relationships between LOINC parts, terms, and concepts.^18,19  As such, LOINC does not fulfill the definition of a computable medical terminology as articulated by Cimino et al,²⁰  that being one based in concept orientation, concept definition, and polyhierarchy. These characteristics enable the logical inference between concepts, parts, and part elements.

The LOINC terminology is built by using a construct of 6 independent parts/axes that define the laboratory test performed and reported. Each axis provides a range of granularity values that can be used to define a particular LOINC term, which can be expounded by the data provided in additional database fields, including order and/or observation classification, “related names 2,” and “long common name” fields.⁶  The “related names 2” field of the LOINC database serves to assist the user in identifying a concept's common names, abbreviations, previous nomenclature, common misspellings, and associated diseases, which might help the user select one LOINC code over another. However, this study identified inaccuracies in the “related names 2” field in the case of LOINC codes specifying D-dimer as the component, which calls into question the curation practices used for this field. Specifically, hyphenation appears to have resulted in the erroneous association of the concepts “Rh” or “Rhesus,” representing the D- Rh blood group commonly denoted “D-,” and “Dextro,” representing a dextrorotatory stereoisomer usually denoted “D-,” with the concept of “D-dimer.”

The LOINC database does not provide a mechanism within the LOINC part structure to support data aggregation based on general concepts subsuming more granular concepts. A computable hierarchy would enable such data aggregation and promote interoperability. Without such mechanisms, interoperability is impeded.²¹  Therefore, interoperability and subsequent data analytics activities are limited to alignment/equivalence of each part. A pertinent example within the public health domain is influenza testing. For example, 513 LOINC terms exist for influenza testing where the component being tested is some form of influenza. It is necessary for public health surveillance within and between health jurisdictions to identify the frequency of influenza testing and results.¹¹  One such reason may simply be an increase in influenza testing by any method (RNA based, antigen based, or other methods) in order to identify potential outbreaks early. Methods vary by testing location, laboratory, and public health jurisdiction. Regardless of methods and specificity to a particular strain of influenza, the ability to aggregate testing method frequency and their results is important in this use case.¹¹  In the current form of LOINC, a comprehensive enumeration of all LOINC terms must be assembled and maintained in order to address this use case because LOINC does not currently have a grouping mechanism for “like” concepts. Logical relationships between LOINC part elements would improve the terminology's utility but cannot address all aggregation issues. Definition of LOINC parts and subsequent LOINC code selection extend LOINC coding issues.

The sophistication of LOINC code selection extends beyond the more simplistic challenges that have already been highlighted. For many laboratory tests, such as cardiac troponins (cTn), variability in the generation of the assay can be associated with critical changes in the test sensitivity, which in turn may be tightly linked with the way the test results are used clinically. High-sensitivity troponin assays, characterized by an imprecision measured as the coefficient of variation below 10% at the 99th percentile value, measure cTn concentrations levels below the 99th percentile in more than 50% of normal individuals, which is 5- to 100-fold lower than conventional assays.^22–24  There are LOINC codes that enable the distinction between conventional assays (LOINC 10839-9) and HscTn assays (LOINC 49563-0) via the method column by indicating that the detection limit is 0.01 ng/mL or lower (equivalent to 10.0 ng/L or 0.01 μg/L). The CAP CRT survey's troponin I method master list distinguishes between HscTn assays and conventional assays (see Data Supplement S2), and provides different method codes for each high-sensitivity assay, as some manufacturer platforms can accommodate both conventional and HscTn assays. As demonstrated by our results, assigning the appropriate LOINC codes for these assays even when the method codes are known is not straightforward and is prone to error. All participant laboratories that reported using a high-sensitivity assay chose the LOINC code for the conventional assay, while none of the laboratories that selected the high-sensitivity LOINC code reported using a high-sensitivity assay.

Prior studies examining the accuracy of LOINC mapping found that choosing different ‘method,' ‘scale,' and ‘property' attributes were the most common reasons for incongruent code choices among 3 large institutions.²¹  Similar to the current study, the authors found variations in the utilization of LOINC that limit true interoperability of laboratory data across different institutions. The authors identified issues with internal and external consistency within the LOINC system that need to be corrected and recommended the addition of computable relationships between the terms to allow for reliable inferences about the data to improve semantic interoperability.²¹  International efforts to address this shortcoming include the cooperative agreement between the Regenstrief Institute²⁵  and SNOMED International.²⁶  While promising, these efforts require ongoing support and cooperation of each standards-developing organization.

Laboratory directors today are challenged with how to respond to increasing institutional demands to accommodate the aggregation of laboratory data from multiple testing sources into single displays, tables, or charts. This might include data from outreach laboratories, in-office laboratories, and point-of-care testing devices. The thought is that the ability to display trends over time may allow more rapid assimilation of the data by the treating clinicians. LOINC codes would be used to identify the “same test” performed at multiple locations. Inappropriate assignment of LOINC codes can therefore lead to misinterpretation of data and data trends, potentially leading to inappropriate treatment decisions; furthermore, even for tests that have been correctly assigned the same LOINC code, it does not mean that the data are interchangeable or aggregable.¹⁷  If LOINC codes are used as the crux for interfacing laboratory results from one system to another, an inappropriately selected LOINC code could result in data being displayed with an incorrect name, thereby giving health care providers incorrect information.

There are several limitations to this study. This analysis was based on a limited sample size, since only 5% of laboratories responded to the survey. However, although the data set is relatively small, the volume and breadth of errors identified in these common laboratory tests were significant. The survey did not ask about the background (eg, laboratory staff or employees in information technology) and qualifications of the personnel who assign LOINC codes within the laboratory or if the personnel had received special training on LOINC code selection best practices. LOINC states on its Web site that it does not provide certification to any organization or individual.²⁷  Given the complexity of LOINC coding, its requirement for use under United States federal law, and the lack of any requirement for LOINC coders to have any certification, this may be a significant factor in the lack of standardized coding between laboratories, seen in this study. Laboratories were asked whether they sent their LOINC codes to the EHR, but they were not asked if their EHRs actually used the LOINC code in any way to determine where to put the results in its database. Since many laboratories simply pass the LOINC code to the EHR without using the LOINC code itself to determine where the result should file, the rate of the use of the LOINC code for interoperability is unknown.

The data in this study should raise concern about the safety, reliability, and utility of using LOINC codes for clinical research or to aggregate data, particularly in evolving areas, for example, protocol development to rule-in or rule-out acute myocardial infarction using data from HscTn assays, as many errors were uncovered in this small sample size alone. This study provides insight into the real-world selection and usage of LOINC codes, and the limitations of the current system as it operates across a variety of laboratory settings.