Background/Purpose: Avascular necrosis (AVN) of bone is a painful, disabling condition. Studies aimed at improving the diagnosis or treatment of AVN require accurate case-finding methods. We examined the sensitivity, specificity, positive predictive value (PPV) and positive likelihood ratio (LR⁺) of alternative algorithms that use claims data to identify cases of AVN of the upper and lower extremities.

Methods: Using a centralized clinical data registry from a large academic hospital, we identified all adults aged ≥18 years who underwent MRI of an upper and/or lower extremity joint for any indication between January 1, 2010 and June 1, 2011. We examined the performance characteristics (sensitivity, specificity, PPV, and LR⁺) of four algorithms (A – D) using International Classification of Diseases, 9th edition (ICD-9) codes for AVN (ICD-9, 733.4X) (Table). The algorithms ranged from least stringent (Algorithm A, requiring ≥1 1 ICD-9 code) to most stringent (Algorithm D, requiring ≥ 3 ICD-9 codes at least 30 days apart). Only ICD-9 codes within 6 months of MRI diagnosis were included. We compared cases identified by each algorithm to the gold standard of a clinical MRI reading by a radiologist confirming “avascular necrosis” or “osteonecrosis.” We calculated 95% confidence intervals (CI) using the normal approximation of the binomial distribution. 

Results: A total of 11,878 patients who underwent MRI of the upper and lower extremities during the 1.5 year period were included in this study.  The prevalence of AVN using the gold standard of MRI was 0.7%, with 83 total cases of AVN. Algorithm A had a sensitivity of 81.9% (95% CI 71.9-89.5), with a PPV of 48.6% (95%CI 40.0-57.2) and a LR⁺ of 134 (95% CI 104-173). The PPV of algorithm D increased to 61.4% (95% CI 47.6-74.0) with a LR⁺ of 226 (95% CI 139-368), although the sensitivity decreased to 42.2% (95% CI 31.4-53.5) (Table). The specificity of all four algorithms ranged from 99.0 to 99.8%.

Conclusion: In this study, we demonstrated that the PPV for AVN among patients who underwent MRI ranged from 49-61% in different ICD-9 code-based algorithms. Given its high sensitivity, Algorithm A (requiring at least 1 ICD-9 code for AVN) appears best suited for situations in which it would be problematic to miss AVN cases, and confirming cases to exclude false positives with further chart review is feasible.  Algorithm B, requiring ≥2 ICD-9 codes at least 7 days apart, had the highest PPV and might be recommended when further validation is not feasible, although misclassification may occur. These algorithms provide an efficient way to identify AVN cases in administrative data, and the PPVs will be greater in populations with higher disease prevalence such as SLE or orthopedic cohorts. Of note, since all patients in this study underwent MRI, cases of asymptomatic or mild AVN that did not prompt MRI evaluation would not be detected with these methods.

Table. Performance characteristics of ICD-9 code algorithms for the diagnosis of avascular necrosis (AVN) of bone
Algorithms	Sensitivity (95% CI),%	Specificity (95% CI), %	PPV (95% CI),%	LR+ (95% CI)
A: ≥1 ICD-9 code for AVN	81.9 (71.9-89.5)	99.4 (99.2-99.5)	48.6 (40.0-57.2)	134.2 (104.4-172.6)
B: ≥2 ICD-9 codes for AVN at least 7 days apart	56.6 (45.3-67.5)	99.8 (99.7-99.8)	62.7 (50.7-73.6)	238.5 (157.5-361.3)
C: ≥2 ICD-9 codes for AVN at least 30 days apart	42.2 (31.4-53.5)	99.8 (99.7-99.9)	60.3 (46.6-80.0)	216.3 (133.9- 349.4)
D: ≥3 ICD-9 codes for AVN at least 30 days apart	42.2 (31.4-53.5)	99.8 (99.7-99.9)	61.4 (47.6-74.0)	226.1 (138.8-368.2)

---

« Back to 2014 ACR/ARHP Annual Meeting

ACR Meeting Abstracts - https://acrabstracts.org/abstract/validation-of-the-diagnosis-of-avascular-necrosis-of-bone-in-administrative-data/