ADR-DQPU: A Novel ADR Signal Detection Using Deep Reinforcement and Positive-Unlabeled Learning

The medical community has grappled with the challenge of analysis and early detection of severe and unknown adverse drug reactions (ADRs) from Spontaneous Reporting Systems (SRSs) like the FDA Adverse Event Reporting System (FAERS), which often lack professional verification and have inherent uncertainties. These limitations have exacerbated the difficulty of training a robust machine-learning model for detecting ADR signals from SRSs. A solution is to use some authoritative knowledge bases of ADRs, such as SIDER and BioSNAP, which contain limited confirmed ADR relationships (positive), resulting in a relatively small training set compared to the substantial amount of unknown data (unlabeled). This paper proposes a novel ADR signal detection method, ADR-DQPU, to alleviate the issues above by integrating deep reinforcement Q-learning and positive-unlabeled learning. Upon validation using FAERS data, our model outperformed six traditional methods, exhibiting an overall accuracy improvement of 26.45%, an average accuracy improvement of 52.15%, a precision enhancement of 1.89%, a recall improvement of 18.57%, and an F1 score improvement of 10.95%. In comparison to two state-of-the-art machine learning methods, our approach demonstrated an overall accuracy improvement of 64.1%, an average accuracy improvement of 28.23%, a slight decrease of 1.91% in precision, a recall improvement of 55.56%, and an F1 score improvement of 45.53%.