ToxicFreeMed: BERT-enhanced BPR algorithm for toxic-free medication recommendation

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Pharmacology

Background:

• Critically-ill patients often require multiple medications (polypharmacy), increasing the risk of adverse drug-drug interactions (DDIs).
• Accurate and safe medication recommendations are crucial for managing complex conditions and improving patient outcomes.
• Existing recommendation systems may not adequately balance therapeutic accuracy with pharmacological safety.

Purpose of the Study:

• To develop a novel model, ToxicFreeMed, for providing toxic-free medication recommendations.
• To integrate patient electronic discharge notes, drug descriptions, and DDI knowledge graphs for enhanced recommendation accuracy and safety.
• To reduce the incidence of DDIs in medication recommendations for critically-ill patients.

Main Methods:

• Utilized a BERT-enhanced Bayesian Personalized Ranking (BPR) algorithm.
• Integrated pretrained embeddings from patient discharge notes and drug descriptions.
• Employed a multitask learning framework to optimize both ranking accuracy and a toxicity-weighted DDI loss.
• Evaluated the model on the MIMIC-III dataset, augmented with DrugBank knowledge.

Main Results:

• ToxicFreeMed effectively balances recommendation accuracy and drug safety.
• The model outperformed existing strong recommendation algorithms in accuracy and DDI reduction.
• Demonstrated a significant reduction in potential drug-drug interactions.

Conclusions:

• ToxicFreeMed is a promising approach for assisting clinicians in identifying safer and more effective medication options.
• The model has the potential to improve treatment strategies for patients with complex conditions.
• Highlights the value of integrating diverse data sources for personalized and safe medication recommendations.