Development and validation of case-finding algorithms to identify prosthetic joint infections after total knee arthroplasty in Veterans Health Administration data

Abstract

Purpose:

To determine the positive predictive values (PPVs) of ICD-9, ICD-10, and current procedural terminology (CPT)-based diagnostic coding algorithms to identify prosthetic joint infection (PJI) following knee arthroplasty (TKA) within the United States Veterans Health Administration.

Methods:

We identified patients with: (1) hospital discharge ICD-9 or ICD-10 diagnosis of PJI, (2) ICD-9, ICD-10, or CPT procedure code for TKA prior to PJI diagnosis, (3) CPT code for knee X-ray within ±90 days of the PJI diagnosis, and (4) at least 1 CPT code for arthrocentesis, arthrotomy, blood culture, or microbiologic procedure within ±90 days of the PJI diagnosis date. Separate samples of patients identified with the ICD-9 and ICD-10-based PJI diagnoses were obtained, stratified by TKA procedure volume at each medical center. Medical records of sampled patients were reviewed by infectious disease clinicians to adjudicate PJI events. The PPV (95% confidence interval [CI]) for the ICD-9 and ICD-10 PJI algorithms were calculated.

Results:

Among a sample of 80 patients meeting the ICD-9 PJI algorithm, 60 (PPV 75.0%, [CI 64.1%–84.0%]) had confirmed PJI. Among 80 patients who met the ICD-10 PJI algorithm, 68 (PPV 85.0%, [CI 75.3%–92.0%]) had a confirmed diagnosis.

Conclusions:

An algorithm consisting of an ICD-9 or ICD-10 PJI diagnosis following a TKA code combined with CPT codes for a knee X-ray and either a relevant surgical procedure or microbiologic culture yielded a PPV of 75.0% (ICD-9) and 85.0% (ICD-10), for confirmed PJI events and could be considered for use in future pharmacoepidemiologic studies.

1. INTRODUCTION

Total knee arthroplasties (TKAs) are one of the most common elective surgeries in the United States (US) and are highly successful in treating pain and disability from osteoarthritis or other knee conditions.¹ Prosthetic joint infection (PJI) is an infrequent but serious adverse outcome, resulting in high morbidity for patients and a heavy economic burden on the health care system.^2–4 Despite the significance of PJIs, there have been few population-representative studies evaluating the incidence, determinants, and outcomes of these infections. Electronic healthcare databases could serve as valuable sources of information to study the epidemiology of PJIs following TKA, allowing for potentially large, representative, and well-characterized study samples with longitudinal follow-up after joint replacement. However, in order to study the epidemiology of PJIs, methods to identify these events accurately within the data sources must be developed and validated.

Validated case-finding algorithms to identify PJIs after joint arthroplasty in electronic health record (EHR) and U.S. databases are lacking. PJI has been validated following total hip replacements in the Danish Hip Arthroplasty Register, but registries contain prospectively entered clinical data, with uniformly completed data fields and may differ from administrative or EHR data.⁵ Recently, a validation study of PJI case-finding algorithms following hip or knee arthroplasty in Canadian administrative data using International Classification of Diseases, 10th Revision (ICD-10) codes found that diagnosis codes for PJI alone performed reasonably well with a positive predictive value (PPV) of 78% (95% confidence interval [CI], 74%–82%).⁶ However, the transportability of these algorithms to other settings, is unknown. Reliable methods to ascertain PJIs in US EHR databases would permit the comprehensive study of the pharmacoepidemiology of antimicrobials both for prevention and treatment of these infections.

To address this methodologic need and to identify case-finding algorithms for PJI within Veterans Health Administration (VHA) data, we developed and assessed the performance characteristics of three pre-specified approaches using International Classification of Diseases, Ninth Revision (ICD-9) or ICD-10 diagnoses, alone or in combination with relevant current procedural terminology (CPT) codes.

2. METHODS

2.1. Design and data source

We conducted a retrospective study between January 1, 2000 and October 31, 2020 using EHR data available from the VHA. The VHA is the largest integrated health care system in the United States,⁷ providing care to over 5.8 million persons at 1255 health care facilities nationwide,⁸ including 170 hospitals or medical centers and 1074 outpatient clinics.⁹ All care contacts, including inpatient and outpatient visits, are recorded daily in the VA’s EHR system. Data are available from ambulatory, Emergency Department, and hospital encounters, including demographics, medical diagnoses, procedures, laboratory values, microbiology and pathology reports, dispensed medications, admission and progress notes, operative notes and hospital discharge summaries. The study was approved by the Institutional Review Boards of the Corporal Michael J. Crescenz Philadelphia VA Medical Center and the University of Pennsylvania with a waiver of informed consent.

2.2. Patients selected for validation

We evaluated three separate approaches to identify PJI events within VHA data. To permit use in both the ICD-9 and ICD-10 eras, we separately developed an ICD-9-based and ICD-10-based algorithm for each approach, so that a total of six algorithms were examined. For each algorithm, we utilized a stratified sampling method, whereby a random sample of 80 potential PJI events was taken, and stratified by annual TKA volume at each site to ensure representation of charts from very low (<10 TKAs per year), low (10–19 TKAs per year), medium (20–99 TKAs per year) and high (≥100 TKAs) volume VHA centers.

Our first approach identified potential PJI events based on ICD diagnosis codes alone. For these algorithms, we identified Veterans who had: Criterion (i) a hospital ICD-9 (Algorithm 1a) or ICD-10 (Algorithm 1b) discharge diagnosis (in principal or contributory position) of PJI (see Table 1 for list of diagnoses) recorded between January 1, 2000 and October 31, 2020; and Criterion (ii) TKA ICD-9 (81.54) or ICD-10 (OSRD0xx, OSRC0xx) code prior to the PJI diagnosis. We evaluated all hospital discharge PJI diagnoses, regardless of whether recorded in primary or secondary position, because we sought to identify PJI events that not only occurred in the outpatient setting and prompted hospital admission, but also those that developed within the hospital, such as those which might occur in patients admitted with bacteremia who develop an infected TKA from secondary seeding after hospitalization. An algorithm for TKA consisting of ICD-9 and CPT codes has previously been shown to identify this procedure accurately within VHA data, with a PPV of 95.2% (95% CI 92%–99%).¹⁰

TABLE 1.

Prosthetic joint infection-related diagnosis codes and descriptions

Code type	Code number	Code description
ICD-9 and ICD-10 diagnoses for prosthetic joint infection
ICD-9	996.60	Infection and inflammatory reaction due to internal prosthetic device implant and graft
	996.66	Infection and inflammatory reaction due to unspecified internal joint prosthesis
	996.67	Infection and inflammatory reaction due to other internal orthopedic device, implant, and graft
	996.69	Infection and inflammatory reaction due to other internal prosthetic device, implant, and graft
ICD-10	T81.42XA	Infection following a procedure, deep incisional surgical site, initial encounter
	T81.43XA	Infection following a procedure, organ and space surgical site, initial encounter
	T84.50XA	Infection and inflammatory reaction due to unspecified internal joint prosthesis, initial encounter
	T84.53XA	Infection and inflammatory reaction due to internal right knee prosthesis, initial encounter
	T84.54XA	Infection and inflammatory reaction due to internal left knee prosthesis, initial encounter
	T84.59XA	Infection and inflammatory reaction due to other internal joint prosthesis, initial encounter
	T84.7XXA	Infection and inflammatory reaction due to other internal orthopedic prosthetic devices, implants and grafts, initial encounter
	T85.79XA	Infection due to other internal prosthetic devices, implants and grafts, initial encounter
Current procedural terminology codes for knee X-ray
CPT	73560	X-ray knee 1 or 2 views
	73562	X-ray knee 3 views
	73564	X-ray knee 4 views
	73565	X-ray knee bilateral standing
Current procedural terminology codes for relevant surgical procedures or microbiological culture
CPT	20610	Arthrocentesis, aspiration and/or injection; major joint
	20611	Arthrocentesis, aspiration and/or injection, major joint, ultrasound guided
	27310	Arthrotomy, knee, with exploration, drainage, or removal of foreign body (e.g., infection)
	27330	Arthrotomy, knee; with synovial biopsy only
	27331	Arthrotomy, knee; including joint exploration, biopsy, or removal of loose or foreign bodies
	27334	Arthrotomy, with synovectomy, knee; anterior or posterior
	27335	Arthrotomy, with synovectomy, knee; anterior and posterior including popliteal area
	87040	Culture, bacterial; blood, with isolation and presumptive identification of isolates
	87999	Unlisted microbiology procedure

Abbreviations: CPT, current procedural terminology; ICD-9, International Classification of Diseases, 9th revision; ICD-10, International Classification of Diseases, 10th Revision.

To attempt to enhance the accuracy of our algorithms, our second approach identified potential PJI events based on Criteria (i)–(ii) plus the addition of: Criterion (iii) CPT code for a knee X-ray (Table 1) ±90 days from the PJI diagnosis date. The rationale for Criterion iii was to increase the likelihood that a PJI was attributed to the TKA, given plain radiographs are performed universally in all cases of suspected PJI and would help eliminate detection of PJIs of other joints as well as other infections of prosthetic materials.

Finally, our third approach identified possible PJIs building on Criteria (i)–(iii) plus the addition of: Criterion (iv) CPT code for arthrocentesis of a major joint, arthrotomy of a knee, blood culture, or microbiologic procedure (Table 1) ±90 days from the PJI diagnosis date. The addition of Criterion (iv) was aimed at providing further criteria to support the likelihood of a PJI and to decrease the detection of superficial surgical site infections that never proceeded to invasive tissue sampling of any kind (without which it is impossible to meet definite or probable PJI diagnostic criteria [discussed below]). The microbiologic procedure code included was for “microbiology procedure not otherwise specified,” since this was commonly recorded among patients with hospital PJI diagnoses and likely corresponded to microbiology testing of synovial, tissue, bone, or other relevant cultures of the involved knee or possibly other sites. These criteria are depicted in Figure 1.

FIGURE 1.

Design diagram highlighting criteria for selection of veterans for validation. ICD, International Classification of Diseases; PJI, prosthetic joint infection; TKA, total knee arthroplasty

2.3. Prosthetic joint infection definition

A PJI event was confirmed if the patient met the clinical case definition for definite or probable PJI. A definite PJI diagnosis was based on the 2013 International Consensus Group definition for PJI, which was adapted from the Musculoskeletal Infection Society definition of this infection (Table 2).^11,12 In addition, we identified a probable PJI diagnosis if: (1) at least two minor criteria were present, and (2) prescription for an antibiotic at the time of the PJI diagnoses was recorded and continued for a duration of at least of 4 weeks. The probable PJI diagnosis was included since criteria needed to classify a PJI diagnosis as definite may not have been present because: (1) requisite laboratory results were absent (e.g., patient was never tested or refused the test; procedure needed to obtain the laboratory result was unsuccessful), or (2) empiric antimicrobial therapy prevented diagnosis by microbiologic culture. Moreover, prior to the 2011 American Academy of Orthopedic Surgeons Clinical Practice Guidelines for the Diagnosis of PJI of the Hips and Knees, which recommended that multiple microbiologic periprosthetic cultures be obtained intraoperatively in patients with suspected PJI, it was common for only one culture to be collected.¹³ Consequently, meeting criteria for a definite PJI diagnosis might be more difficult before this date. For these reasons, we chose to evaluate our algorithms compared to confirmed definite or probable PJI diagnoses.

TABLE 2.

Criteria for defining definite and probable prosthetic joint infections

Definite prosthetic joint infection	Probable prosthetic joint infection
One of the following criteria present during admission or ≤30 days prior:	Two or more of the following present during admission or ≤30 days prior:
• Two positive periprosthetic cultures with phenotypically identical organisms	• Serum C-reactive protein of ≥10 mg/L (or ≥ 1.0 mg/dL) AND erythrocyte sedimentation rate ≥30 mm/h
• A sinus tract communicating with the joint	• Synovial fluid white blood cell count of ≥3000 cells/μL OR +or ++ change on leukocyte esterase test strip
OR
Three or more the following findings present during admission or ≤30 days prior:	• Synovial fluid polymorphonuclear neutrophil percentage of ≥80%
	• >5 neutrophils per high power field in five high power fields (×400) or acute inflammation reported on pathology
• Serum C-reactive protein of ≥10 mg/L (or ≥1.0 mg/dL) AND erythrocyte sedimentation rate ≥30 mm/h
	• A single positive synovial fluid culture
• Synovial fluid white blood cell count of ≥3000 cells/μL OR + or ++ change on leukocyte esterase test strip	AND
	• Prescription for antibiotic at time of PJI event, for minimum of 4-week course
• Synovial fluid polymorphonuclear neutrophil percentage of ≥80%
• >5 neutrophils per high power field in five high power fields (×400) or acute inflammation reported on pathology
• A single positive synovial fluid culture

2.4. Confirmation of outcomes

VHA electronic medical record data were abstracted onto structured forms (Appendix S1) using Research Electronic Data Capture (REDCap) tools hosted by the VA.^14,15 Forms collected information from the following components of the VHA electronic medical record: (1) clinician hospital admission notes, progress notes, and discharge summaries (to confirm presence of a PJI event after TKA, presence of a sinus tract, and antibiotic duration); (2) operative reports (to confirm presence of a sinus tract); (3) laboratory results (to collect results of C-reactive protein (CRP), erythrocyte sedimentation rate (ESR), synovial fluid leukocyte count, percentage of synovial fluid leukocytes that were polymorphonuclear cells (PMN %), peripheral blood cultures, synovial fluid microbiologic cultures, and tissue cultures); and (4) surgical pathology reports (to evaluate for presence of acute inflammation within bone or synovium).

Abstraction forms were independently reviewed by two infectious diseases clinicians (B.L., T.Y., or E.J.W.) to determine if the criteria for PJI occurred during a hospitalization (classified as definite, probable, or no event). Disagreement resulted in review by a third infectious diseases clinician to arbitrate the event.

2.5. Statistical analyses

We calculated the PPV with 95% CI of each ICD-9 and ICD-10-based algorithm for confirmed definite or probable PJI. Our focus was on PPV since a sufficiently high PPV provides confidence that identified outcomes are true events. Since probable events are very likely to be PJI diagnoses only with one or more missing lab values and were treated clinically as PJIs with a prolonged antimicrobial course (Table 2), we felt that including these in the calculation for PPV was appropriate. We estimated that a sample of 80 patients for each algorithm would allow determination of the PPV with a sufficiently narrow 95% CI width of ±10%, assuming a PPV of 80%. We also calculated the percent observed agreement between the two adjudicators, using a quadratic weighting system to rate disagreements, based on an inverse-square spacing, whereby a weight of 0.75 for agreement was used for any disagreements between definite and probable cases, but a weight of 0.0 for agreement was used for disagreements between probable and no event.¹⁶ All data were analyzed using Stata 16.1 (Stata Corp, College Station, TX).

3. RESULTS

Of the 127 862 TKAs identified in the VHA from 2000 to 2020, 4511 (3.52%) went on to meet criteria for Algorithms 1a or 1b, 4002 (3.13%) for Algorithms 2a or 2b and finally, 2861 (2.24%) for Algorithms 3a or 3b. For each of the six algorithms, 80 patients meeting the algorithm were sampled. Four hundred twenty-five unique PJI events total were included, as some were sampled for more than one algorithm. Appendix S2 summarizes the baseline demographics of these populations. PJI events sampled spanned the entire inclusion period from 2000 to 2020 (Appendix S3), and there was representation from VHA sites with varying TKA volumes (Appendix S4). Medical records from all sampled patients were available and all sampled cases were successfully adjudicated.

We found that Algorithms 1a and 1b had low PPV (Table 3). Among the 80 patients sampled to evaluate Algorithm 1a, 43 (PPV, 53.8% [95% CI, 42.2%–65.0%]) were confirmed to have a PJI event (42 definite, 1 probable). Similarly, for the 80 patients sampled to evaluate Algorithm 1b, 48 (PPV, 60.0% [95% CI, 48.4%–70.8%]) were confirmed to have PJI (46 definite, 2 probable). The percent agreement among adjudicators for these algorithms was 99.8%.

TABLE 3.

Positive predictive values (with 95% confidence intervals) of case-identifying algorithms to identify prosthetic joint infection among patients with total knee arthroplasty

				Reason not confirmed
					Other infectious processes				Non-infectious processes
Algorithm	ICD version	N confirmed PJI out of 80	PPVa (95% CI)	Did not meet criteria	Skin/soft tissue	PJI, other joint	Other ortho	Non-ortho hardware	Other complication	Second stage of two-step revision
1	9	43	53.8% (42.2%–65.0%)	17	9	0	2	3	6	0
	10	48	60.0% (48.4%–70.8%)	11	7	4	5	2	1	2
2	9	56	70.0% (58.7%–79.7%)	9	9	0	2	0	4	0
	10	53	66.3% (54.8%–76.4%)	11	8	1	0	0	4	2
3	9	60	75.0% (64.1%–84.0%)	10	3	0	1	1	3	2
	10	68	85.0% (75.3%–92.0%)	7	3	0	2	0	0	7

Abbreviations: CI, confidence interval; ICD, International Classification of Diseases; Non-ortho hardware, non-orthopedic prosthetic infection such as abdominal mesh infection; Other ortho, other orthopedic non-arthroplasty infection such as infected nail or screw; PJI, prosthetic joint infection; PPV, positive predictive value.

^aPositive predictive value based on cases adjudicated as definite or probable events.

Algorithms 2a and 2b (addition of knee X-ray within 90 days of structured codes) demonstrated only modest improvement in performance (Table 3). For Algorithm 2a, among 80 patients sampled, 56 (PPV, 70.0% [95% CI, 58.7%–79.7%]) were confirmed to have a PJI (50 definite, 6 probable). For algorithm 2b, 53 of the 80 patients (PPV, 66.3% [95% CI, 54.8%–76.4%]) had a PJI (53 definite, 0 probable). The percent agreement between adjudicators was 100%.

In contrast, Algorithms 3a and 3b demonstrated the highest PPVs. Of the 80 patients sampled to evaluate Algorithm 3a, 60 (PPV, 75.0% [95% CI, 64.1%–84.0%) had a confirmed PJI (57 definite, 3 probable). Of the 80 patients sampled for Algorithm 3b, 68 (PPV, 85.0% [95% CI, 75.3%–92.0%]) had PJI (65 definite, 3 probable). The percent agreement among adjudicators for these algorithms was 98.91%.

Table 3 reports the most common reasons that each algorithm was not adjudicated as a definite or probable PJI diagnosis. The most commons reasons included insufficient laboratory criteria, presence of superficial skin/soft tissue infection but no PJI, or presence of another orthopedic or non-orthopedic infection. Additional clinical details about the probable PJI diagnoses are included in Appendix S5.

4. DISCUSSION

In this study, we evaluated the PPVs of three ICD-9-based and three ICD-10-based coding algorithms to identify PJI diagnoses within VHA data. Algorithms 1a and 1b, which used ICD diagnosis codes for PJI preceded by a TKA procedure code, performed poorly, with PPVs ≤60%. Algorithms 2a and 2b additionally included a knee X-ray within ±90 days of the PJI diagnosis date but had only modestly improved performance, with PPVs ranging from 66.3% (ICD-10) to 70.0% (ICD-9). Algorithms 3a and 3b added the requirement of a procedure code for an arthrocentesis, arthrotomy, or microbiology procedure to the parameters of Algorithms 2a and 2b and performed the best, achieving PPVs of 75.0% (ICD-9) and 85.0% (ICD-10).

Algorithms 1a and 1b were likely inadequate because of the lack of specificity of the PJI diagnosis codes. Both ICD-9 and ICD-10 classification systems include PJI diagnosis codes that have broad descriptions, such as “infection or inflammatory reaction due to unspecified internal joint prosthesis” and “infection due to other internal prosthetic devices, implants, and grafts.” As a result, Algorithms 1a and 1b captured patients not only with PJIs of the knee, but also individuals with TKAs who had prosthetic hip or shoulder infections, local inflammatory reactions, and other non-TKA-related infections (Table 3). Algorithms 2a and 2b, which added the requirement of a knee X-ray within 90 days of the PJI diagnosis, reduced capture of non-TKA-related infections, but still identified a number of patients who had superficial skin and soft tissue infections overlying their TKA in the absence of definite or probable PJI criteria.

To enhance the ability to detect PJI after TKA, we developed a third set of algorithms that additionally required at least one CPT code for an arthrocentesis of a major joint, arthrotomy of a knee, blood culture, or microbiologic culture not otherwise specified and found that these algorithms (Algorithms 3a and 3b) yielded PPVs of ≥75.0%. The addition of this criterion to our algorithms led to substantially fewer superficial skin/soft tissue infections being identified, selecting for patients who had synovial fluid or tissue sampling and higher concern for deep infection. ICD-10-based Algorithm 3b had a higher PPV (85.0%) than the ICD-9-based Algorithm 3a (75.0%). This may have been due to the greater specificity of the ICD-10 diagnoses (Table 1) and the inclusion of qualifiers such as “initial encounter” (vs. “subsequent encounter”) in the ICD-10 code descriptions.

Our selection of the additional criteria within Algorithms 3a and 3b was based on several considerations. While microbiology results are available within VHA data, given the substantial proportion of PJIs which may be culture-negative (range, 5%–45%),¹⁷ we did not require a positive microbiologic culture. Similarly, we opted not to include criteria for antibiotic treatment, as there are a number of patients who get transferred outside of the VHA for surgical management and antibiotic treatment. These reasons all support our inclusion of codes centered around the diagnosis of PJI, rather than treatment.

Our results contrast with those of a recent validation study by Kandel et al. using diagnostic and procedure codes to identify PJI following total hip or knee arthroplasties within Canadian administrative data, where an ICD-10 discharge diagnosis of PJI alone had a PPV of 78% (95% CI 74%–82%).⁶ Possible explanations for their improved performance of PJI diagnosis codes alone, compared to our Algorithms 1a and 1b, could be due to their inclusion of both hip and knee PJI diagnoses as well as differences in local diagnostic and coding practices.

Our study has a number of strengths. We used rigorous clinical case definitions against which to classify definite and probable PJIs.^11,12 We required clinical criteria for case definitions to be present 30 days prior to or during the PJI hospital admission, which provides reasonable confidence in the date of PJI diagnosis. We used an innovative approach to overcome the lack of specificity in the ICD-9/−10 diagnosis codes, requiring an X-ray of the knee to capture TKA PJIs as well as one or more CPT codes for either a joint sampling procedure or microbiologic culture. Finally, we evaluated separately the PPVs of ICD-9-based and ICD-10-based coding algorithms, to permit their use for studies in different eras.¹⁸ An algorithm using ICD-9 era PJI diagnosis following total joint arthroplasty has not yet been validated, so this is novel.

This study has several potential limitations. First, misclassification of PJI events could have occurred during adjudication, but we minimized this possibility by using a pre-specified, standardized definition for PJI and employing two infectious diseases-trained adjudicators to confirm events, with a third to arbitrate outcomes in cases of disagreement. While our algorithms accurately identified PJI diagnosis dates, due to the clinical and heterogeneous natural history of PJI, the date of onset of PJI is not identified by this algorithm. Second, the algorithms with the highest PPVs (Algorithms 3a and 3b) might miss some PJI events, since they require a hospital discharge diagnosis of PJI, antecedent TKA code, knee X-ray CPT code, and one or more CPT code(s) for surgical procedures or microbiologic cultures. Third, we did not determine the negative predictive value (NPV), sensitivity, or specificity of the algorithms. Given the low frequency of PJI events, the negative predictive value is expected to be high.¹⁹ Calculating the sensitivity of the algorithm was not feasible, as there is no longitudinal PJI registry within the VHA with which to identify all the individuals with a known PJI in the health system. Finally, this study assessed the validity of coding algorithms within the VHA, but the accuracy of these algorithms might differ in other data sources and so should be evaluated prior to use.

In conclusion, we developed three ICD-9-based and three ICD-10-based case-finding algorithms to identify PJIs following TKA within national VHA data. We found that algorithms consisting of an ICD-9 or ICD-10 PJI hospital discharge diagnosis, antecedent TKA code, and CPT codes for a knee X-ray plus a related surgical procedure (arthrocentesis or arthrotomy) or microbiologic culture identified a PJI diagnosis with ≥75% PPV. These algorithms could be used in future epidemiologic studies to evaluate important determinants of PJI following TKA within VHA data.

Key Points.

• Electronic healthcare databases could be valuable resources to examine risk of prosthetic joint infection (PJI) following total knee arthroplasty (TKA), but algorithms to identify PJI after TKA have not been developed or validated within U.S. healthcare databases.

• We developed separate algorithms based upon International Classification of Diseases, Ninth Revision (ICD-9) and International Classification of Diseases, Tenth Revision (ICD-10) diagnoses for prosthetic joint infection, in combination with current procedural terminology (CPT) codes, to identify PJI events following TKA within US Veterans Health Administration data.

• We identified patients with: (i) hospital discharge ICD-9 or ICD-10 diagnosis for PJI, (ii) ICD-9, ICD-10 or CPT procedure code for TKA prior to PJI diagnosis, (iii) CPT code for knee X-ray within ±90 days of the PJI diagnosis, and (iv) at least 1 CPT code for arthrocentesis, arthrotomy, blood culture, or microbiologic procedure within ±90 days of the PJI diagnosis date.

• An algorithm consisting of an ICD-9 or ICD-10 code for PJI following a procedure code for TKA, combined with CPT codes for a knee X-ray and any one of arthrocentesis, arthrotomy, blood culture, or microbiologic procedure within ±90 days of the PJI diagnosis date, had a PPV of 75.0% (ICD-9) and 85.0% (ICD-10), respectively, for confirmed PJI.

• These algorithms could be considered for use in future studies of PJI following TKA within Veterans Health Administration data.