人工智能辅助上市后药品不良反应自动化识别与流程重构OA

Objective:This study aims to develop and validate an artificial intelligence(AI)screening toolkit and an associated management workflow for post-marketing,multi-source free-text data,with the aim of improving the automated identification of adverse drug reaction(ADR)and the processing efficiency within pharmacovigilance(PV).Methods:Differentiated modeling was implemented according to text characteristics from distinct sources.A text convolutional neural network(text-CNN)model was used for short texts such as hotline call records and sales visit records,while a Transformer-based model was applied to long texts such as academic literature.Automatic screening was enhanced by integrating expanded lexicons for medicinal products and ADR terms,and a human-AI collaborative workflow redesign was proposed.Results:Across three scenarios,namely hotline call records,sales visit records,and academic literature,the model achieved a recall of 99％in all cases.The corresponding precision values were 75.5％,53.5％,and 53.0％,respectively,and the accuracy was 97.2％,99.0％,and 75.6％,respectively.Conclusions:A differentiated AI strategy for multi-source unstructured text substantially enhances the sensitivity of ADR detection and screening efficiency.The supporting human-AI collaboration and system integration enable proactive surveillance and upstream process shifts,reducing the risk of underreporting and strengthening compliance.