Interpretable one-class classification framework for prescription error detection using BERT embeddings and...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Pharmacy

Background:

• Accurate prescriptions and medication administration are vital for patient safety and clinical efficacy.
• Prescription errors lead to increased healthcare costs and adverse events.
• Existing error detection methods (rule-based, supervised ML) have limitations in adaptability and data requirements.

Purpose of the Study:

• To develop and evaluate a novel prescription error detection method using a one-class classification approach.
• To overcome limitations of existing methods, particularly the need for labeled error data.
• To provide interpretable insights into model predictions for enhanced clinical trust.

Main Methods:

• Utilized the MIMIC database for a large-scale prescription dataset.
• Employed advanced language modeling (BERT embeddings) and dimensionality reduction (Principal Component Analysis).
• Implemented a one-class classification model (Local Outlier Factor) for anomaly detection, enhanced with LIME and SHAP for explainability.

Main Results:

• The proposed method effectively detects potential prescription errors without requiring labeled error data.
• Achieved high performance metrics: Precision=81.71%, Recall=87.32%, F1-score=86.84%.
• Explainability methods (LIME, SHAP) provided clinicians with interpretable insights, increasing trust.

Conclusions:

• The one-class classification approach offers a robust and adaptable solution for prescription error detection.
• This method significantly enhances patient safety and can reduce healthcare-associated costs.
• The integration of explainable AI fosters trust and facilitates clinical adoption of automated error detection systems.