Identifying Alzheimer's disease and related disorders via diagnostic codes in Veterans with heart failure

Abstract: Cognitive impairment due to Alzheimer's disease and related disorders (ADRD) threatens self-management ability when co-occurring with heart failure. ADRD also increases the incremental cost of heart failure care in managed Medicare organizations. ICD codes facilitate the identification of persons with ADRD for population health management and observational research. However, multiple ICD-based approaches have developed, and the differences are poorly understood. This study aims to determine the inter-rater reliability between the VA Dementia ICD Code List and the Chronic Conditions Warehouse (CCW) ADRD algorithm when applied to a cohort of Veterans hospitalized with heart failure. This cross-sectional study used secondary data on a sample of 373,897 Veterans hospitalized in VA medical centers between October 1, 2011, and September 30, 2020, with a primary admission diagnosis of heart failure (see Supplemental Methods). We randomly selected one index admission in Veterans with more than one eligible admission to allow sampling across the clinical course of heart failure. The Median (IQR) number of heart failure admissions per eligible Veteran was 2 (1–3). All study procedures were approved by the IRB at the VA Providence Healthcare System. We applied the VA and CCW algorithms using the same 3-year reference period. We computed Cohen's kappa coefficient with a 95% confidence interval according to methods previously described.4 We used Pearson's chi-square test to assess independence between algorithm and race. Overall, the cohort included 373 97 Veterans, of whom 364,341 (97%) were male and 74,478 (20%) were Black (Table 1). The CCW code list classified 107,690 (29%) as having ADRD and 266,207 (71%) as not. The VA code list classified 61,796 (17%) as having ADRD and 312,101 (84%) as not. The VA algorithm did not classify any Veterans as having ADRD who were not so classified by the CCW code list. The number of Veterans classified as having ADRD by both the CCW and VA code lists was 61,796 (17%); CCW but not VA, 45894 (12%); neither CCW nor VA, 266207 (71%). The raw percent agreement was 88%. Cohen's kappa coefficient (95% CI) for interrater reliability between the CCW and VA code lists was 0.66 (0.65–0.66), indicating substantial agreement.5 For race, the chi-square test produced a p-value of <0.001, inconsistent with the null hypothesis that algorithm performance is independent of race. The VA algorithm identified a higher proportion of Black Veterans than the CCW algorithm. In this study of Veterans with heart failure, ICD-based algorithms estimated ADRD prevalences of 17% and 29%, consistent with the pooled results of 32 prior studies finding a percentage (95% CI) of 20% (13%–28%).6 The inter-rater reliability between the CCW and VA-developed ADRD algorithms was substantially greater than would be expected by random chance. However, the two algorithms classified nearly 1 in 8 Veterans differently. The difference in classification was completely unidirectional: the VA-developed algorithm identified solely Veterans that CCW also identified, whereas CCW classified 74% more Veterans as having ADRD. The composition of these algorithms accounts for their differing patterns of classification. The CCW algorithm differs from the VA algorithm by including several non-specific codes with loose equivalence or association to ADRD. VA-only codes, while specific to a narrowly interpreted concept of ADRD, did not measurably alter the algorithm's performance in this sample of Veterans because there were no Veterans with claims containing these codes. The race data demonstrate a small but significant difference in the racial composition between the Veterans identified with the VA and CCW algorithms compared to the CCW algorithm alone, a finding that merits further investigation. This work has several important limitations. We did not compare the coding algorithms to other sources of diagnostic information; therefore, this work does not address the potential problem of underdiagnosis or misdiagnosis of ADRD. This work also does not address the diagnosis of mild cognitive impairment. Our cohort contained only Veterans hospitalized for heart failure, so our observations may not be generalizable to other populations. In Veterans with heart failure, the CCW ADRD algorithm identifies more cases than the VA algorithm due to its inclusion of nonspecific codes with loose equivalence to the ADRD concept. Investigators and health systems desiring tight adherence to the ADRD concept should consider using the VA algorithm. Investigators and health systems who wish to obtain a larger sample inclusive of ADRD and several loosely associated diagnoses should consider using the CCW algorithm.