Jiang Yongxian, Zhao Shan, Li Zhangyue, Li Gen, Qin Bo, Yi Zhanmiao
Objective To analyze the current research status and application value of artificial intelligence (AI) technology in adverse drug event prediction. Methods Literature on the application of AI technology in drug adverse event prediction from databases such as PubMed, Embase, Cochrane Library, China Biomedical Literature Service System, and CNKI was searched. The extracted literature data included the first author, publication year, source country, research type, research field, adverse event type, drugs, modeling algorithm, performance indicators, etc. Evidence levels were assessed using the 2011 edition of the Oxford Center for Evidence-Based Medicine (a total of 5 levels, with level 1 being the highest and level 5 being the lowest). Descriptive statistical analysis was performed using Excel 2021. Results A total of 52 articles were ultimately included, including 10 prospective studies and 42 retrospective studies. The evidence quality assessment of the literature revealed that 10 studies were non-randomized controlled trials or prospective cohort studies (evidence of level 2), 41 were retrospective cohort studies or case-control studies (evidence of level 3), and 1 was a single-center longitudinal cohort study and cross-sectional study (evidence of level 4). Fifty-two articles used a total of 10 AI models, of which 23 articles only used 1 model, and 29 articles used 2 or more models, totaling 111 model applications. The more frequently used models were random forest model, neural network model, logistic regression model, reinforcement model, support vector machine model, Bayesian network model, and long short-term memory (LSTM), accounting for 22.52% (25/111), 18.02% (20/111), 17.12% (19/111), 11.71% (13/111), 10.81% (12/111), 6.31% (7/111), and 5.41% (6/111), respectively. The model with best performance was LSTM, with an area under the receiver operating characteristic curve (ROC-AUC) of 0.85±0.04, accuracy of 0.76±0.08, recall rate of 0.83±0.11, specificity of 0.74±0.05, F1 score of 0.79±0.18, and precision of 0.86±0.11. More articles were related to adverse events related to cardiovascular, digestive, and immune systems, accounting for 17.31% (9/52), 11.54% (6/52), and 9.62% (5/52), respectively. The random forest model showed superior performance in predicting adverse events of immunosuppressants, with an ROC-AUC of 0.82, recall rate of 0.57, and precision of 0.80. The neural network model showed better performance in predicting adverse events of antitumor drugs, with an ROC-AUC of 0.87, recall rate of 0.87, and precision of 0.73. The logistic regression model performed well in predicting adverse events of immunosuppressants, with an ROC-AUC of 0.84, recall rate of 0.57, and precision of 0.81. From the perspective of systems involved in adverse events, the reinforcement model demonstrated the best performance for predicting adverse events in the cardiovascular system (ROC-AUC=0.82, precision=0.62); for the digestive system, the random forest model achieved the highest performance (ROC-AUC=0.80, recall=0.59, precision=0.92); for the immune system, the neural network model performed the best (ROC- AUC: 0.90±0.06, recall: 0.86±0.12, precision: 0.93±0.09). Conclusion The performance of existing AI models in predicting adverse drug events varies, with optimal models differing across systems. Current research primarily focuses on predicting adverse events in the cardiovascular, digestive, and immune systems, but clinical practicality remains insufficient.
