A probabilistic model for reducing medication errors: A sensitivity analysis using Electronic Health Records data.

Area of Science:

• Health Informatics
• Clinical Decision Support Systems
• Medical Record Analysis

Background:

• Medication errors pose a significant threat to healthcare quality.
• Clinical decision support systems (CDSS) are crucial for mitigating these errors.
• Existing CDSS importance is often underestimated, necessitating robust validation.

Purpose of the Study:

• To validate an existing probabilistic model, AESOP, for automated medication error identification.
• To perform a sensitivity analysis of the AESOP model using electronic medical record data.
• To assess the model's performance across different healthcare settings and specialties.

Main Methods:

• Constructed a knowledge base of 2.22 million disease-medication and 0.78 million medication-medication associations from Taiwan claims data (2009-2011).
• Utilized 0.6 million outpatient prescriptions for model validation.
• Conducted a sensitivity analysis with 11 thresholds (α = [0.5; 1.5]) and compared 2400 prescriptions against a physician-adjudicated gold standard.

Main Results:

• The AESOP model achieved high accuracy (over 80%) and positive predictive value (over 85%) in identifying medication errors.
• Sensitivity ranged from 80-96%, while negative predictive values varied (45-75%) across cardiology, neurology, and ophthalmology departments.
• Model performance was evaluated across 121 prescription results at various thresholds.

Conclusions:

• Sensitivity analysis confirmed the AESOP model's effectiveness in identifying medication errors across different hospitals.
• Optimal threshold selection requires balancing false positives and false negatives.
• Model performance is influenced by clinical specialty and intended application, highlighting the need for tailored implementation.