Predicting clinically significant medication errors: Development and validation of a risk stratification model using incident reporting data
TL;DR
A parsimonious predictive model using readily available incident report data can effectively identify medication errors at elevated risk of clinical significance in a tertiary care setting.
Abstract
Background Medication error reporting systems capture large volumes of incidents, yet only a small proportion lead to clinically significant patient outcomes. Distinguishing these high-risk errors from the majority of low-impact events remains a major challenge for healthcare systems. This study aimed to develop a predictive model to identify medication errors at elevated risk of clinical significance in a tertiary care setting. Methods A retrospective cross-sectional study was conducted at Al-Noor Specialist Hospital in Makkah, Saudi Arabia, analyzing medication error reports from January 2023 to August 2025. Clinically significant errors were defined as those reaching the patient with or without harm (NCC MERP Categories C–G). Univariate analyses identified candidate predictors, followed by multivariable logistic regression with category collapsing for model efficiency. Model performance was assessed using area under the receiver operating characteristic curve (AUC) and Brier score. Risk stratification divided predicted probabilities into low, moderate, and high-risk groups. Results Among 6,397 medication errors, 227 (3.5%) were clinically significant. The final model demonstrated good discrimination (AUC = 0.825) and calibration (Brier score = 0.029). Independent predictors of clinically significant errors were high-risk error stages including administration, dispensing, and monitoring (OR = 19.1, 95% CI: 14.1–26.1), high-risk medication classes such as antibiotics and oncology agents (OR = 2.70, 95% CI: 1.62–4.78), intermediate or other medication classes (OR = 2.34, 95% CI: 1.39–4.18), selection or information errors (OR = 2.23, 95% CI: 1.40–3.46), and lack of staff experience (OR = 1.82, 95% CI: 1.33–2.49). Variables describing who detected or who committed the error were deliberately excluded from the model because they are not available at the moment an error occurs. Risk stratification classified 10.1% of errors as high-risk, capturing 62.6% of all clinically significant events. Conclusion A parsimonious predictive model using readily available incident report data can effectively identify medication errors at elevated risk of clinical significance. The risk-stratification framework enables prioritization of safety resources toward high-risk incidents requiring immediate intervention.