Prediction of Prednisolone Dose Correction Using Machine Learning.

Area of Science:

• Pharmacovigilance
• Health Informatics
• Machine Learning

Background:

• Prescription errors, particularly incorrect dosing, pose significant risks to patient safety, especially with high-alert medications like oral corticosteroids.
• Oral prednisolone, a commonly prescribed corticosteroid, is susceptible to dose-related prescription errors.

Purpose of the Study:

• To develop and validate machine learning (ML) models for predicting dose-related prescription modifications of oral prednisolone.
• To address the challenge of highly imbalanced data inherent in identifying prescription errors.

Main Methods:

• Prescription data for oral prednisolone tablets were extracted from electronic medical records.
• Cluster analysis was used to group clinical departments by prescription patterns.
• Synthetic Minority Over-sampling Technique (SMOTE) was applied for data preprocessing.
• Five ML models (SVM, KNN, GB, RF, BRF) and logistic regression were trained and evaluated using stratified cross-validation and a holdout test set.

Main Results:

• The Balanced Random Forest (BRF) model achieved high performance (ROC-AUC: 0.917, recall: 0.951) on the original imbalanced dataset.
• SMOTE preprocessing improved performance across most models.
• With SMOTE, the Support Vector Machine (SVM) model showed the highest ROC-AUC (0.820) and BRF demonstrated strong recall (0.634) on the test set.

Conclusions:

• Machine learning, particularly using techniques like SMOTE and BRF, can effectively predict dose-related prescription modifications for oral prednisolone.
• These models are valuable tools for complex dose audits and improving medication safety in clinical practice.