This study determined that ICD-9 codes 433.X1, 434.XX, and V12.54 had positive predictive values >90%, and may be used to focus care on stroke patients.
To determine the positive predictive values of inpatient and outpatient ICD-9 codes and status code V12.54 for identifying confirmed history of stroke or transient ischemic attack (cerebral event) among patients within a managed care organization.
Retrospective, cohort study.
Inpatient hospital claims and outpatient visit records were used to identify patients with ICD-9 codes (430.XX to 438.XX) or status code V12.54 in the primary or secondary position recorded between January 1, 2001, and December 31, 2009. A standardized chart abstraction tool was used by trained chart abstractors blinded to the coding to confirm the cerebral event and classify stroke type. Positive predictive values (PPVs) were calculated for each code based on care setting.
A total of 4689 patients with 10,376 unique stroke codes recorded in the administrative data were reviewed. Of these, 2785 (59.4%) patients had a confirmed cerebral event. The codes with PPV less than 90% were 434.XX, 433 .X1, and V12.54 where codes were recorded in both the inpatient and outpatient settings. Overall, inpatient-only codes produced higher PPVs; however, relatively fewer events were captured in this setting.
Administrative ICD-9 codes 434.XX, 433.X1, and V12.54 had consistently high PPVs in identifying patients with a confirmed cerebral event. These codes could be used as part of a probabilistic approach to focus care activities on patients with the highest likelihood of a cerebral event.
Am J Manag Care. 2014;20(2):e27-e34
Stroke is the fourth-leading cause of death, the most common neurologic disorder, and a leading cause of long-term disability in the United States.1 Despite widely available national clinical practice treatment guidelines for ischemic stroke, studies demonstrate that a significant proportion of these patients have uncontrolled modifiable risk factors and remain undertreated.2-4 Barriers to implementing treatment guidelines include patient, provider, and healthcare system factors. Developing validated disease registries may allow better tracking and management of patients with stroke.
Administrative data, including International Classification of Diseases, Ninth Revision (ICD-9) diagnosis codes, can be a powerful tool for identifying eligible patients for disease state registries. Healthcare organizations have developed stroke registries to address treatment gaps and improve quality of care, and support surveillance, epidemiologic and economic research for patients with ischemic stroke.4-8 However, cerebrovascular-related ICD-9 diagnosis codes may not reliably identify patients with a confirmed stroke.4-7,9-17
Studies evaluating the accuracy of cerebrovascular ICD-9 codes have relied primarily on patients discharged following emergency department (ED) visits or hospitalizations.6,8,9-11,13-16 These data may not be available for patients transitioning to a new healthcare system. Patients new to a healthcare system and/or patients who were never treated in the ED or hospital could be seen in the outpatient setting where either self-report or clinical diagnosis of cerebrovascular disease is made. Thus, incorporating diagnoses and health status administrative codes from the outpatient setting may improve the ability to identify patients with a confirmed stroke or transient ischemic attack (TIA). The purpose of this study was to determine the positive predictive value (PPV) of cerebrovascular ICD-9 codes using data from both inpatient and outpatient care settings to capture patients, both incident and prevalent, with cerebral events including stroke or TIA. The information from this study may be useful for healthcare systems to develop and/or improve stroke registries.
Study Design and Setting
This retrospective study was conducted at Kaiser Permanente Colorado (KPCO), an integrated healthcare delivery system providing services to more than 500,000 members at 24 medical offices in the Denver-Boulder metropolitan area. At KPCO, all providers utilize an electronic medical record (EMR), in which all office visit, vital sign, imaging, laboratory, hospital discharge summaries, pharmacy, etc, data are housed. An EMR has been used at KPCO since 1998. In addition to the provider interface, the EMR has a data repository that can be queried electronically to extract relevant data. The majority of KPCO patients receive prescription medications from in-house pharmacies for a co-payment. This study was approved by the KPCO Institutional Review Board with a waiver of informed consent.
All patients at least 18 years of age with at least 1 inpatient stay or outpatient medical office visit with an ICD-9 diagnosis code between 430.xx and 438.xx or status code V12.54 (Personal History of Stroke) (Appendix) in the primary or secondary position recorded between January 1, 2001, and December 31, 2009, were identified administratively from inpatient hospital claim databases and outpatient visit records from the EMR data repository. Medical information (eg, provider visit notes, hospital discharge summaries) in the EMR was available from January 1, 1998, onward. In some cases, scanned medical information was available prior to 1998.
While there is some disagreement for using all diagnosis codes versus limiting to primary and secondary positions,12,17 we used codes from both positions to maximize the yield of patients with which to examine for true strokes. Patients were categorized by the first 3 digits of the identified diagnosis code (eg, a patient with a diagnosis of 433.10 would be categorized as 433) or the status code alone (ie, V12.54) and the setting type where the code was recorded (inpatient only, outpatient only, or the same code recorded in both care settings). All of a patient’s cerebrovascular codes identified in the administrative data during the study period were captured; thus, a patient could have had more than 1 code category assigned in the inpatient and/or outpatient settings. There were no prespecified exclusion criteria.
Information from each eligible patient’s EMR was considered the standard for determination of a confirmed stroke or TIA. An abstraction tool using elements from the Rochester Minnesota Stroke study form18 was developed with input from KPCO neurologists. The tool was used to identify patients with a confirmed incident or prevalent stroke or TIA (cerebral event), identify the cerebral event date, and, if a confirmed stroke was found, classify the stroke type (ischemic, hemorrhagic, or unknown). For purposes of this manuscript, all cerebral events (including TIAs) are referred to as a stroke unless specifically describing/discussing TIAs. Identifying a confirmed stroke required sufficient clinical and/or radiologic evidence of diagnosis (focal neurological symptoms and/or computed tomography [CT] scan/magnetic resonance imaging [MRI] verifying ischemia, infarct, or hemorrhage). Recognizing the difficulty in objectively diagnosing TIA, a confirmed TIA was defined as a coded diagnosis with written confirmation by physician(s) in hospital progress/discharge note(s) or medical office note(s).
Given that secondary prevention management of TIA is similar to that of ischemic stroke,2 TIA was included with the ischemic stroke classification. Intracerebral and subarachnoid bleeding was classified as hemorrhagic stroke. The “unknown” classification was used when chart review revealed clinical evidence of stroke, but insufficient clinical or radiologic evidence was available to establish stroke type. Trained clinical pharmacy specialist reviewers used the abstraction tool and all available information in the EMR to determine whether patients had confirmed stroke and to classify stroke type. Reviewers were blinded to administrative ICD-9 code(s). A kappa score was calculated for a random sample of 20 patients with the result of 0.8 indicating a high degree of inter-rater reliability for stroke confirmation.17 Additionally, the study investigators met with the reviewers to review the discrepancies and agree on how to code certain cases moving forward.
Prescriptions for antiplatelet medications (ie, ticlopidine [Ticlid], clopidogrel [Plavix], dipyridamole [Persantine], and dipyridamole/ ASA [Aggrenox]), use of imaging (CT and/or MRI scans), and outpatient neurology department visits within 180 days of initial cerebrovascular ICD-9 code(s) were identified using queries of KPCO-integrated electronic databases. These stroke-related health service use indicators were used in conjunction with the study ICD-9 codes to determine if the PPVs of administrative data to identify strokes could be improved.
Positive predictive values for each ICD-9 code were determined by dividing the number of confirmed stroke events by the total number of events recorded in the administrative data for the specified code. The primary study outcome was the PPV with 95% confidence intervals (CIs) for correctly identifying confirmed cerebral events.19 PPVs were calculated for each ICD-9 code in the setting where the code was recorded (ie, inpatient setting only, outpatient setting only, or the same code recorded in both settings). Secondary outcomes included assessments of PPV for identifying confirmed hemorrhagic and ischemic strokes, respectively for each hemorrhagic (430, 431, 432) and ischemic (433, 434, 435, 436, 437, 438, V12.54) code. Positive predictive values for these outcomes were determined by dividing the number of specific confirmed events (ischemic or hemorrhagic) by the total number of events recorded in the administrative data for the specified code. Subanalyses were performed to determine the robustness of PPV estimates under various scenarios.
The impact of using study ICD-9 codes in conjunction with health service indicators was determined by calculating PPVs for each ICD-9 code individually and in combination with various health service indicators. To accomplish this, only patients who were exposed to the health service indicator and had the ICD-9 code diagnosis were used to calculate the PPV. To assess the impact of ICD-9 code position (primary or secondary), secondary position inpatient codes and patients with only a secondary position inpatient code were removed and PPVs recalculated. To assess the impact of including TIA, all patients with confirmed TIA were removed and PPVs recalculated. To assess the impact of the ICD-9 code 436 not being inclusive of stroke, because the code excludes “cerebrovascular accident (CVA) NOS, Stroke” as of October 1, 2004, the PPVs of code 436 recorded before and on/after this date were recalculated. To assess the impact of the use of codes with specific “infarction” terms, codes 433.XX and 434.XX were categorized individually by codes that contain the infarction term (ie, 433.X1 and 434.X1) or not and the PPVs were recalculated. Positive predictive values, were determined with SAS version 9.1.3 (SAS Institute Inc, Cary, North Carolina) using Proc Freq with weighting by the count of patients having a specific diagnosis and the exact binomial function to determine 95% CIs.
Patient characteristics were reported as means with standard deviations for interval-level characteristics. These characteristics were assessed for distribution normality and appropriate tests (eg, t test, rank-sum test) were used to assess differences between groups. To assess differences in proportions between groups on dichotomous characteristics, Pearson’s x2 test of association was utilized. A 2-sided alpha level was set at <.05.
A total of 4689 patients with 10,376 unique study administrative ICD-9 codes were reviewed. Of these, 2785 (59.4%) patients had a cerebral event confirmed by EMR review. The majority of patients had ICD-9 codes from the outpatient setting (82.6%) while 1.3% and 16.1% were from inpatient and both outpatient and inpatient settings, respectively (). The most commonly identified cerebral event types were non-cardioembolic strokes (34.8%) and TIAs (31.1%). Cerebral event type was unknown in 15.4% of cases. Patients with confirmed cerebral events had a higher mean count of unique ICD-9 codes, were slightly older, more likely to have purchased a prescription antiplatelet drug, and more likely to have had CT or MRI imaging. Positive predictive values for “intracerebral hemorrhage”( 431), “acute but ill-defined cerebrovascular disease” (436), and “personal history of stroke” (V12.54) were greater than 90% when recorded in both the inpatient and outpatient settings but identified small numbers of patients; thus, associated 95% CIs were wide (). “Occlusion of cerebral arteries” (434) recorded in either the inpatient-only or in both the inpatient and outpatient settings also achieved PPVs greater than 90% and identified large numbers of patients. Codes 434 and V12.54 in the outpatient-only setting identified the most patients and had reasonably high PPVs (both 89%). “Other and unspecified intracranial hemorrhage” (432) and “other and ill-defined cerebrovascular disease” (437) performed poorly regardless of setting (PPVs <50%). Overall, codes recorded in both the inpatient and outpatient settings yielded the highest PPVs but identified fewer patients compared with those recorded only in the inpatient or outpatient settings. Hemorrhagic stroke codes identified fewer patients than ischemic stroke codes (). Hemorrhagic stroke codes recorded only in the inpatient setting had higher PPVs than outpatient-only codes. Overall, ischemic codes tended to have higher PPVs than hemorrhagic codes (). Approximately 15% of code 436 patients had their code recorded on/after October 1, 2004, and the vast majority of these (99%) were in the outpatient setting. Nevertheless, there was no appreciable change in PPV before and after 436’s coding modification. Code 433.X1 with specific mention of infarction did increase the PPV appreciably but only approximately 8% of patients with a 433.XX code had a 433.X1 code. Code 434.X1 with specific mention of infarction did not alter the PPV appreciably over presence of the code 434.XX (Table 2).
In general, inclusion of health service indicators and exclusion of secondary inpatient diagnoses and TIA patients identified fewer patients, widened 95% CIs and did not change the PPVs substantially. Overall, inclusion of diagnostic imaging did not change PPV estimates, nor did inclusion of neurology visits. Inclusion of prescription antiplatelet exposure slightly improved PPVs (mean difference in PPVs across care settings = 1.5 [± 4.2], median difference in PPVs across care settings = 1.0). However, for codes 434 and V12.54, inclusion of prescription antiplatelet exposure resulted in more marked improvement in PPV (mean = 8.6 [± 9.7], median = 4). Removing secondary position inpatient codes did not affect the PPVs appreciably (mean = —0.5 [± 3.0] and median = 0), but removing patients with a confirmed TIA (code 435) slightly reduced PPVs overall (mean = –2.0 [± 2.5] and median = –1).
Disease registries provide opportunities for health systems to improve management of patients with chronic disease states. Initial patient identification using coded administrative data is an important part of developing a validated patient registry, particularly if the codes have high PPV. Using a standardized abstraction tool adapted from the Rochester Minnesota Stroke study form,18 we found only 60% of patients identified from administrative data using cerebrovascular ICD-9 codes had a confirmed cerebral event. We found that the settings where ICD-9 codes were recorded influenced both the accuracy of diagnosis and yield of identified cases. Codes recorded in both inpatient and outpatient settings had higher absolute PPVs, but identified fewer patients than codes recorded in only 1 of these settings. Attempts to improve the accuracy of ICD-9 codes through various combinations with health services indicators produced, at best, only moderate improvements, with the exception of combining purchases of prescription antiplatelets with codes 434 and V12.54. The incorporation of the setting where the ICD-9 codes were recorded and using combinations health service indicators are unique aspects of our study. Nevertheless, in most cases, our PPV estimates were associated with wide 95% CIs despite a relatively large sample size suggesting that administratively coded data elements may lack sufficient accuracy to be relied on without confirmation of cerebral events via medical record review.
Several inpatient studies have evaluated the accuracy of ICD-9 codes for identifying ischemic stroke by assessing the sensitivity, specificity, and/or PPV of ICD-9 diagnosis codes 430 to 438.4-7,9-16 These studies also demonstrated less than optimal accuracy of ICD-9 codes in identifying confirmed stroke patients. One study reported a PPV of only 47% for ICD-9 codes between 430 and 438 for correctly identifying incident stroke events.5 Additionally, several studies have revealed that registries derived from hospital discharge codes overestimate stroke.5,6,16 While our use of ICD-9 codes recorded in inpatient and/or outpatient settings appears to modestly improve the accuracy of identifying confirmed cerebral events, the use of ICD-9 codes alone appears to lead to a high percentage of false-positive diagnoses, such that about 40% of events identified with these commonly used ICD-9 codes are not confirmed cerebral events.6
Benesch and colleagues found that limiting inpatient ICD-9 codes to those listed in the primary discharge position increased stroke PPV.10 We hoped to increase event capture by using ICD-9 codes recorded in the primary or secondary discharge positions. Nevertheless, we also performed a subanalysis using only the primary position and found that removing secondary position codes did not decrease our inpatient PPVs. Our results may have differed from their study since we had relatively low rates of false positive strokes.
We reported PPV estimates for codes recorded only in the outpatient setting, capturing patients who may or may not have been hospitalized for treatment prior to enrolling in our health plan. We found that considerably more patients were identified in the outpatient setting. Similar to prior studies, we found that ICD-9 codes 434 and 436 had high PPVs for patients with confirmed ischemic stroke.10,11 We were able to show that the PPVs for these codes slightly improved when the codes were recorded in both the inpatient and outpatient settings (97% and 93%, respectively). Since October 1, 2004, when 436 coding changes were implemented, the number of patients with this code decreased considerably, making the utility of this code to identify patients with stroke less robust.
We assessed ICD-9 code 435 (TIA) because treatment guidelines are similar for patients who experienced an ischemic stroke and TIA is an important risk factor for stroke, with 90-day risks reported as high as 17%.2 Interestingly, we found that code 435 was one of the most prevalent codes, capturing 1329 patients who had their cerebral event confirmed, and its PPVs were relatively high across settings (>84%) indicating potential utility for this code if identification of patients with probable TIA is desirable.
Historically, ICD-9 codes for hemorrhagic stroke have reported higher PPVs than ischemic stroke codes.14,15 However, we found that, overall, hemorrhagic stroke codes had lower PPVs than ischemic stroke codes, perhaps because there were fewer patients with hemorrhagic stroke codes. We did find that inpatient hemorrhagic codes had higher PPVs than outpatient hemorrhagic codes suggesting that inpatient hemorrhagic events are coded more accurately.
It is not surprising that ICD-9 codes 432 “other and nonspecified intracranial hemorrhage” and 437 “other and illdefined cerebrovascular disease” performed poorly based on their definitions. For example 437 includes a variety of conditions under the umbrella of “ill defined,” including cerebral atherosclerosis, chronic cerebral ischemia, hypertensive encephalopathy, non-ruptured cerebral aneurysm, moyamoya disease, nonpyogenic thrombosis of intracranial venous sinus, and transient global amnesia.10
Overall, the inclusion of health service indicators such as diagnostic imaging and neurology visits did not change appreciably the PPV estimates, nor did evaluating code 434.X1 with specific infarction terms. Code 433.X1 with specific mention of infarction did increase the PPV appreciably but identified only a small number of patients. Inclusion of prescription antiplatelet exposure slightly improved PPVs. We identified no other studies that employed health service indicators to assess their effects on PPVs.
The ideal cerebrovascular ICD-9 code would capture a large number of patients and maintain a high PPV. We found that codes 434 and V12.54 came closest to ideal when recorded in the outpatient setting. Code 434.XX is likely highly predictive because the operating definition is specific to thrombosis or embolism of cerebral arteries.10 The high predictability of V12.54 is interesting because it is a patient self-reported code. To our knowledge, this is the first study to evaluate the accuracy of ICD-9 status code V12.54, which proved to be highly effective capturing many patients and moderately predictive across the inpatient and outpatient settings. We recommend that health plans wishing to identify administratively patients with a history of stroke or TIA utilize ICD-9 codes 433.X1, 434.XX, 436.XX, and V12.54. These codes demonstrated reasonable predictive abilities across care settings. Conversely, we caution plans from utilizing administratively codes with a low PPV as this may result in a large number of false positives.
Our study was not without limitations. We recognize that PPVs are subject to disease prevalence, however, we were unable to determine sensitivities and specificities as we did not identify or review patients who did not have the specified stroke codes. Given the large number of patients with stroke codes reviewed for this study, the number of non-cases we would have had to review to provide meaningful data would have been too resource intensive. We used a manual chart review process that relied on the accuracy of care and health services use of documentation in the EMR. There may also have been temporal changes in coding practices. The study was conducted at only 1 health plan and coding practices and accuracy may vary across institutions and health plans; thus, our findings may not be applicable at all institutions/ health plans. Similarly, coding and charting practices may have differed between individual physicians, whereby some physicians were more accurate and/or comprehensive.
While we have no evidence of differential coding and charting, we enhanced our chart review process by using a standardized chart abstraction tool to improve consistency and trained chart reviewers who had an inter-rater reliability consistent with a high level of agreement. Additionally, the duration of a member’s enrollment in KPCO may have impacted our ability to confirm strokes. Patients with longer enrollments would have had a greater opportunity to be diagnosed in multiple care settings and/or had medical information to confirm stroke.
The ideal cerebrovascular ICD-9 code would capture a large number of patients and maintain a high PPV. Unfortunately, the majority of cerebrovascular ICD-9 codes lack this ability, making it difficult to accurately identify stroke patients without manual medical chart review. Our results indicate that several stroke codes (433.X1, 434, and V12.54 in particular) had relatively high PPVs across care settings. However, caution should be used with all codes, and they may be best suited to identify patients with a higher likelihood of having had a true stroke. This could provide a means to minimize the number of confirmatory chart reviews that would be needed to accurately identify patients with a true stroke.Author Affiliations: Pharmacy Department, Kaiser Permanente Colorado, Aurora, CO (KLO, MDW, LJL, AMD, JR); University of Colorado Skaggs School of Pharmacy and Pharmaceutical Sciences, Aurora, CO (KLO, TD, JR, AMD); Department of Endocrinology, Colorado Permanente Medical Group and the University of Colorado Health Sciences Center Denver, CO (JAM).
Funding Source: Kaiser Permanente Colorado.
Author Disclosures: The authors (KLO, MDW, TD, LJL, JR, AMD, JAM) report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.
Authorship Information: Concept and design (TD, LJL, JR, AMD); acquisition of data (MDW, TD, JR); analysis and interpretation of data (MDW, TD, LJL, JR, AMD); drafting of the manuscript (MDW, LJL, JR); critical revision of the manuscript for important intellectual content (LJL, JR, AMD); statistical analysis (TD); supervision (JR).
Address correspondence to: Thomas Delate, PhD, 16601 E Centretech Parkway, Aurora, CO 80011. E-mail: firstname.lastname@example.org. Roger VL, Go AS, Lloyd-Jones D, et al. Heart disease and stroke statistics—2012 update: a report from the American Heart Association. Circulation. 2012;125(1):e2-e220.
2. Furie KL, Kasner SE, Adams RJ, et al. Guidelines for the prevention of stroke in patients with stroke or transient ischemic attack: a guideline for healthcare professionals from the American Heart Association/ American Stroke Association. Stroke. 2011;42(1):227-276.
3. Saposnik G, Goodman SG, Leiter LA, et al. Applying the evidence: do patients with stroke, coronary artery disease, or both achieve similar treatment goals? Stroke. 2009;40(4):1417-1424.
4. Asberg S, Henriksson KM, Farahmand B, et al. Ischemic stroke and secondary prevention in clinical practice: a cohort study of 14 529 patients in the Swedish stroke register. Stroke. 2010;41(7):1338-1342.
5. Leibson CL, Naessens JM, Brown RD, Whisnant JP. Accuracy of hospital discharge abstracts for identifying stroke. Stroke. 1994;25(12): 2348-2355.
6. Ellekjaer H, Holmen, Krüger O, Terent, A. Identification of incident stroke in Norway: hospital discharge data compared with a population based stroke register. Stroke. 1999(1);30:56-60.
7. Krarup LH, Boysen G, Janjua H, Prescott E, Truelsen T. Validity of stroke diagnosis in a national registry of patients. Neuroepidemiology. 2007;28(3):150-154.
8. Stegmayr B, Asplund K. Measuring stroke in the population: quality of routine statistics in comparison with a population-based stroke registry. Neuroepidemiology. 1992;11(4-6):204-213.
9. Derby CA, Lapane KL, Feldman HA, Carleton RA. Possible effect of DRGs on the classification of stroke: implications for epidemiological surveillance. Stroke. 2001;32(7):1487-1491.
10. Benesch C, Witter DM Jr, Wilder AL, Duncan PW, Samsa GP, Matchar DB. Inaccuracy of the international classification of diseases (ICD-9) in identifying the diagnosis of ischemic cerebrovascular disease. Neurology. 1997;49(3):660-664.
11. Goldstein LB. Accuracy of ICD-9-CM coding for identification of patients with acute ischemic stroke: effect of modifier codes. Stroke. 1998;29(8):1602-1604.
12. Stineman MG, Granger CV, Hamilton BB, Melvin JL, Fiedler IG. Specificity of ICD-9 CM coding practices for stroke rehabilitation. Am J Phys Med Rehabil. 1993;72(5):318-324.
13. Golomb MR, Garg BP, Saha C, Williams LS. Accuracy and yield of ICD-9 codes for identifying children with ischemic stroke. Neurology. 2006;67(11):2053-2055.
14. Roumie CL, Mitchel E, Gideon PS, Varas-Lorenzo C, Castellsague J, Griffin MR. Validation of ICD-9 codes with a high positive predictive value for incident strokes resulting in hospitalization using Medicaid health data. Pharmacoepidemiol Drug Saf. 2008;17(1):20-26.
15. Tirschwell DL, Longstreth WT Jr. Validating administrative data in stroke research. Stroke. 2002;33(10):2465-2470.
16. Johnsen SP, Overvad K, Sørensen HT, Tjønneland A, Husted SE: Predictive value of stroke and transient ischemic attack discharge diagnoses in The Danish National Registry of Patients. J Clin Epidemiol. 2002;55(6):602-607.
17. Kramer MS, Feinstein AR. Clinical biostatistics: LIV: the biostatistics of concordance. Clin Pharmacol Ther. 1981;29(1):111-123.
18. Petty GW, Brown RD Jr, Whisnant JP, Sicks JD, O’Fallon WM, Wiebers DO. Survival and recurrence after first cerebral infarction: a population-based study in Rochester, Minnesota, 1975 through 1989. Neurology. 1998;50(1):208-216.
19. Fleiss JL, Levin B, Paik MC. Statistical Methods for Rates and Proportions. 3rd ed. Hoboken, NJ: John Wiley & Sons, Inc; 2003