Objectives: Determine sources of error in electronically extracted data from electronic health records.
Study design: Categorical and continuous variables related to early-onset neonatal hypoglycemia were preselected and electronically extracted from records of 100 randomly selected neonates within 3479 births with laboratory-proven early-onset hypoglycemia. Extraction language was written by an information technologist and data validated by blinded manual chart review. Kappa coefficient assessed categorical variables and percent validity continuous variables.
Results: 8/23 (35%) categorical variables had acceptable Κappa (1-0.81); 5/23 (22%) had fair-slight agreement, Κappa < 0.40. Notably, "hypoglycemia" had poor agreement, Κappa 0.16. In contrast, 6/8 continuous variables had validity ≥ 94%. After correcting extraction language, 6/9 variables were corrected and inter-rater validation improved. However, "hypoglycemia" was not corrected, remaining an issue.
Conclusions: Data extraction without validation procedures, especially categorical variables using International Classification of Diseases-9 (ICD-9) codes, often results in incorrect data identification. Electronically extracted data must incorporate built-in validating processes.