Data quality and discharge-time readmission risk in diabetes hospitalisations

BACKGROUND: Electronic health record (EHR) and administrative data are widely reused for readmission modelling, but models may learn from documentation completeness, coding pathways and measurement availability as well as from clinical status. OBJECTIVE: This study assessed how documentation completeness, administrative fields and interpretable clinical predictors shaped discharge-time 30-day readmission risk stratification among hospitalised patients with diabetes. METHOD: A retrospective secondary analysis of the public Diabetes 130-US hospitals for years 1999-2008 dataset was conducted. Encounters with death or hospice-related discharge dispositions were excluded. The outcome was 30-day readmission. Four calibrated logistic regression models compared core clinical variables, administrative fields and missingness indicators under a discharge-time prediction frame. Uncertainty was estimated using patient-level clustered bootstrap resampling. RESULTS: The cohort included 99,343 encounters from 69,990 patient identifiers; 11,314 encounters (11.4%) had 30-day readmission. Weight, maximum glucose serum, A1C result, medical specialty and payer code were missing or not recorded in 96.9%, 94.8%, 83.1%, 48.9% and 39.7% of encounters, respectively. Adding administrative fields increased the area under the receiver operating characteristic curve (AUROC) from 0.640 to 0.667 (ΔAUROC 0.0266, 95% clustered bootstrap confidence interval (CI) 0.0178 to 0.0347). Missingness indicators alone produced little change (ΔAUROC 0.0019, 95% CI -0.0007 to 0.0046). The full model containing administrative fields and missingness indicators showed similar discrimination (AUROC 0.666). CONCLUSION: The modest performance gain was driven mainly by administrative coding and discharge-pathway variables. EHR readmission studies should define prediction timing, report measurement availability and interpret model performance as a product of clinical data generation as well as clinical risk.Implications for health information management practice:Secondary-use EHR modelling should profile variable availability, establish predictor timing and provenance, test administrative-field contributions and report calibrated performance before local use.