We tested two possible strategies for finding incident ON cases in a hypothetical case-control study nested within a large health care cohort. The first would use ICD-9 codes and other administrative data likely to be available to find cases, but cases would not be otherwise verified. A sufficiently high PPV for a given algorithm would suggest that it could be used to identify cases without further need for verification. The second strategy would use the same administrative data to identify potential cases, but the medical records of these potential cases would then be reviewed to confirm which were actual incident cases. This strategy is less efficient than the first but would yield few if any false-positive incident ON cases (high specificity) and could be necessary if no algorithm with a high PPV could be identified. These confirmed cases could be used as the basis of well-designed case-control studies, which can yield valid results comparable to those of cohort studies .
The protocol for this study complied with the Declaration of Helsinki and was approved by the institutional review boards of the Veterans Affairs (VA) Boston Healthcare System, Edith Nourse Rogers Memorial Veterans Hospital, and the Boston University Medical Center. The requirement for informed consent was waived by all of these bodies.
Data source and identification of potential cases
We used data from the VA Boston Healthcare System (VISN1). From 1 October 1998 through 30 September 2006, we attempted to identify veterans seen at VA Boston who were seen for ON; 76,155 veterans were seen in the inpatient or outpatient setting at least once during this period. We used the following ICD-9 codes to identify patients who could have had ON: 733.4× (ON), 732.7 (osteochondritis dissecans), 732.5 (Freiberg infarction), 526.89 (osteoradionecrosis of the jaw), and 526.4 (osteomyelitis or osteitis of the jaw). The second and third codes are conditions related to or easily confused with ON. The fourth and fifth were used as there was no code for ON of the jaw until recently (733.45 since 1 October 2007 ). We called all identified patients 'potential cases' of ON and reviewed all of their medical records. Note that the last digit (x) of the ON code (733.4×) is meant to identify the anatomic site of ON, but this digit was used inconsistently in our database; we therefore did not try to classify site based on the ICD-9 code.
VA has a standardized nationwide medical record system that incorporates almost complete records of all patient visits, including clinic notes, discharge summaries, radiology reports, and actual radiographs. The complete medical record of each veteran can be accessed from any one clinical location, allowing a complete record of their VA-provided health care, even as the patient moves from site to site in the US.
A board-certified rheumatologist (SCV) reviewed all of the available clinical notes, radiology reports, and actual radiographs of each potential case. Using the radiographic criteria established by Sugano and colleagues , we first confirmed that ON was present. Of these criteria, we ignored the requirement for no joint space narrowing or acetabular malformation as some late-stage cases, which clearly started as ON (for example, magnetic resonance imaging band pattern), had progressed to severe OA with attendant narrowing and remodeling of the acetabulum. We also did not use the histologic criterion as histology was usually unavailable. If radiographs were unavailable, a description in the radiology report or clinic notes consistent with these criteria (for example, 'a subchondral lucency') was sufficient to establish the diagnosis.
If ON was present, the case was called 'confirmed' and we examined the clinic notes further to find a description of when the symptoms leading to diagnosis began. If symptoms appeared to begin within 6 months of the first recorded ON code, the case was called 'incident'. If symptoms began more than 6 months prior to the first code or if the time of symptom onset could not be determined, the case was called 'prevalent'.
After initially reviewing the patient records, we decided to use 6 months to define incident ON. We did this for two reasons: (a) sometimes a patient had had symptoms for some months prior to orthopedic referral, at which time a definitive diagnosis was made and the code appeared in the chart; (b) although a diagnosis of ON was sometimes suspected by the referring physician and work-up begun, it was often not coded until the orthopedic consultant saw the patient. Six months appeared to us to offer the best balance between including long-standing cases for which it would be difficult to assess prior risk factors and excluding too many new cases that would otherwise be excluded with a too-stringent definition of time between initial symptoms and codes.
Defining osteonecrosis using administrative data
Because detailed patient data such as those we used for our gold standards are unlikely to be present in large health care databases, we used VA administrative data (that is, data obtained from the administrative records of patient visits as opposed to data extracted directly from the medical records) to test different definitions of ON against the gold standard definitions. Elements from these administrative data included the above codes, dates when codes were assigned, the source of each code (primary care clinic [PCC] versus non-PCC, inpatient visit versus outpatient visit), and the numbers of each code. After the initial review of the records suggested that some ON cases that were first coded around the time of hip replacement had radiographs suggesting long-standing OA as a result of much earlier hip ON, we added definitions including OA. We hoped that this might permit us to identify non-incident ('prevalent') ON and exclude such persons from our search for incident cases. This review also suggested that only the specific ON code (733.4×) was of value in identifying cases; thus, no definition uses any other code for ON (see Results).
Using these elements, we created definitions designed to identify incident ON in administrative databases. These definitions included (a) one or more 733.4× ON codes anywhere in the record (that is, the same definition as that used to identify potential cases), (b) two or more ON codes anywhere in the record, (c) one or more ON codes from an inpatient visit, (d) one or more ON codes from a non-PCC (that is, specialty clinic) visit, (e) (one or more ON codes from an inpatient visit) or (one or more codes from a non-PCC visit), (f) (one or more ON codes from an inpatient visit) or (two or more ON codes, with at least one from a non-PCC visit), (g) (one or more ON codes) and (no prior codes for OA), and (h) (one or more ON codes from an inpatient visit) and (no prior codes for OA).
Testing administrative definitions of osteonecrosis
We first tested each administratively derived definition of ON against the gold standards (strategy 1); that is, we identified all potential cases that would have been found using the administrative definitions, and of these, we found the number that were prevalent cases and the number that were incident cases. In so doing, we found the PPV of each administrative definition for both confirmed (incident + prevalent) and incident (number of cases/number of potential cases using that definition) ON.
We then determined, for each definition, how many cases we could expect to find after reviewing all of the records of the potential cases for that definition (strategy 2). First, we estimated the number of potential cases that would be identified using each administratively defined definition assuming that 1,000 potential cases would be found by the most liberal definition (definition 1). This is 1,000 times the number of potential cases for each definition divided by the number of potential cases for definition 1. Then we found how many of these cases could be expected to be incident ON after reviewing them. This is the number of potential cases for that definition times PPV. Results are expressed as numbers per 1,000 potential cases that would have been found using definition 1.