BACKGROUND: The validity of coding algorithms using diagnostic and/or procedural codes to identify acute kidney injury (AKI), acute liver injury (ALI), severe complications of urinary tract infections (UTI), breast cancer or bladder cancer is not well studied.
OBJECTIVES: To estimate the positive predictive value (PPV) of the coding algorithm for each outcome in a pilot validation study among patients with type 2 diabetes newly initiating anti-diabetic drugs during 2014–2017.
METHODS: We identified provisional cases for each outcome using pre-defined coding algorithms in the HealthCore Integrated Research Database (HIRD). Among provisional cases, we randomly sampled 250 patients hospitalized for AKI, 96 hospitalized for ALI, and 250 who visited the emergency department or had an inpatient hospitalization for severe complications of UTI (i.e., pyelonephritis or urosepsis). We also sampled patients with at least two diagnoses within 60 days apart during outpatient, inpatient, or physician office visits for invasive female breast cancer (n=100) or invasive/in situ bladder cancer (n=20). Two clinicians blinded to information on study drugs, independently reviewed available medical records and adjudicated the case status according to pre-defined clinical criteria. When there was disagreement, final case status was decided by a committee with a third reviewer. PPVs and 95% confidence intervals (CI) of the coding algorithms were estimated.
RESULTS: Among 125 AKI, 45 ALI, and 125 UTI provisional cases reviewed, 48 AKI, 19 ALI, 71 UTI cases were confirmed. After review, 39 AKI, 10 ALI, and 28 UTI cases remained provisional due to insufficient information in selected medical records. Restricting to confirmed cases and non-cases yielded corresponding PPVs of 56% (95%CI, 45%, 66%) for AKI, 54% (95%CI, 38%, 71%) for ALI, and 73% (95%CI, 64%, 82%) for UTI. Among 50 breast and 12 bladder cancer provisional cases reviewed, 41 breast and 9 bladder cancers were confirmed. Restricting to confirmed cases and non-cases yielded corresponding PPVs of 84% (95%CI, 73%, 94%) for breast and 90% (95%CI, 71%, 100%) for bladder cancer after respectively excluding one breast and two bladder cancer provisional cases without sufficient information for adjudication.
CONCLUSIONS: This pilot study suggests that our coding algorithms have high PPVs to capture breast and bladder cancer cases, but lower PPVs for AKI, ALI and severe complications of UTI. Further evaluation of the algorithms for acute outcomes is needed.
