Detecting Potential Adverse Drug Reactions Using a Deep Neural Network Model.

Area of Science:

• Pharmacovigilance and Drug Safety
• Computational Biology and Bioinformatics
• Artificial Intelligence in Medicine

Background:

• Adverse drug reactions (ADRs) are a significant cause of morbidity and mortality, with spontaneous reporting having limitations.
• Current pharmacovigilance methods, primarily spontaneous reporting, suffer from low reporting rates, hindering effective ADR detection.
• There is a critical need for automated methods to enhance the identification and prediction of ADRs.

Purpose of the Study:

• To develop and validate a deep neural network (DNN) model for the automatic detection of potential adverse drug reactions (ADRs).
• To identify potential ADRs for existing drugs and predict ADRs for new drugs using comprehensive drug information.
• To overcome the limitations of spontaneous reporting in pharmacovigilance through an automated approach.

Main Methods:

• A DNN model was designed, integrating chemical, biological, and biomedical drug information.
• Word-embedding techniques processed biomedical literature to represent drug relationships in a vector space.
• A mapping function was developed to accommodate new drugs not present in the initial dataset.

Main Results:

• The DNN model achieved a mean average precision at top-10 of 0.523 and an AUC score of 0.844 for ADR prediction.
• The model accurately predicted ADRs for drugs recorded up to 2012, using data reported up to 2009.
• Experimental validation included 746 existing drugs and 232 new drugs, with 1325 ADRs analyzed.

Conclusions:

• The developed DNN model demonstrates effectiveness in identifying potential ADRs for both known and new drugs.
• The model can predict ADRs even for reactions not previously reported, offering a significant advancement in drug safety monitoring.
• This automated approach enhances pharmacovigilance by providing a more comprehensive and proactive method for ADR detection.