Predicting self-intercepted medication ordering errors using machine learning | JoVE Visualize

Area of Science:

Health Informatics
Clinical Informatics
Medical Error Analysis

Background:

Current methods for analyzing medication ordering errors use small, manually collected samples, limiting scalability and risk factor identification.
Traditional statistical analyses, while informative, require expert guidance and may not capture complex interactions contributing to errors in computerized provider order entry (CPOE) systems.

Purpose of the Study:

To apply machine learning (ML) models to a large dataset of CPOE medication orders to predict erroneous orders and identify contributing factors.
To compare the performance of various ML models against traditional logistic regression for medication error prediction.

Main Methods:

Utilized a dataset of 5,804,192 medication orders, including demographics, clinician characteristics, and order details.
Compared logistic regression, random forest, boosted decision trees, and artificial neural networks, evaluating performance using AUROC and AUPRC metrics.
Identified predictive factors and interacting contexts associated with medication order voids.

Main Results:

Machine learning models demonstrated reasonable accuracy in classifying voided medication orders, with gradient boosted decision trees achieving the highest AUROC (0.7968) and AUPRC (0.0647).
ML models identified predictive factors, such as "student orders," and uncovered complex interacting contexts associated with high error rates, which were missed by previous regression analyses.
Logistic regression performed poorly compared to the ML models.

Conclusions:

Machine learning prediction models utilizing order-entry information show significant promise for enhancing medication error surveillance and patient safety.
The superior performance of models capturing complex interactions highlights the critical role of contextual information in understanding medication ordering errors.
These findings support the use of ML for targeted clinical reviews and improving the overall safety of medication ordering processes.