Using machine learning to predict pharmaceutical interventions during medication prescription review in a hospital...

Area of Science:

• Health Informatics
• Artificial Intelligence in Healthcare
• Clinical Pharmacy

Background:

• Medication errors pose a significant global public health challenge.
• Clinical pharmacists face daily hurdles in reducing inappropriate medication use.
• Advancements in health informatics and AI offer new tools for detecting prescription errors.

Purpose of the Study:

• To compare the efficacy of two machine learning models in predicting the need for pharmaceutical intervention (PI).
• To evaluate the performance of Light Gradient Boosting Machine (LGBM) and Random Forest (RF) models using hospital prescription data.

Main Methods:

• A retrospective study utilized data from a single hospital over four years, encompassing 2,059,847 prescription lines and 260,611 PIs.
• Two tree-based binary classification models, LGBM and RF, were trained and tested on a dataset split into 70% training and 30% testing portions.
• Model performance was assessed on both the global dataset and data stratified by medical care department.

Main Results:

• The LGBM model generally outperformed the RF model across various metrics, including accuracy (86% vs. 85%), precision (80% vs. 42%), specificity (97% vs. 89%), AUC (83% vs. 71%), and F1-score (58% vs. 47%).
• The RF model exhibited higher recall (53% vs. 46%).
• LGBM models trained on the global dataset showed greater effectiveness than those trained on departmental data.

Conclusions:

• The LGBM model demonstrated superior performance in identifying inappropriate prescriptions, suggesting potential improvements in prescription review thoroughness and efficiency.
• This AI-driven approach shows promise for advancing hospital clinical pharmacy.
• Further research is warranted to validate these findings and explore broader applications in optimizing prescription management and enhancing patient care.