(A-267) Using Machine Learning to Predict Classification of Causal and Non-causal Pharmacovigilance Cases

Background: Assessing causality at the product-event case level in pharmacovigilance (PV) is critical yet challenging due to variability in approaches, ranging from expert opinion to reported causality, which often leads to inconsistent documentation. This inconsistency complicates causal inference and could limit the effectiveness of surveillance efforts.

Objectives: To investigate whether machine learning methods can effectively distinguish product event pairs annotated as ‘causal’ or ‘non-causal’ within PV cases, we aimed to develop a binary classifier capable of predicting causality classification with high accuracy and reliability.

Methods: We utilized a dataset comprising 170,400 individual case safety reports (ICSRs) from October 2023 to January 2025, annotated for causality according to PV standards. The data was preprocessed and the subsequent dataset was randomly split into training (75%) and test (25%) sets. Various machine learning models, including Decision Tree, Linear SVM, Random Forest, and XGBoost, were trained using features derived from standard ICSR processing and prior knowledge of predefined safety-designated medical events. Hyperparameter tuning was employed to optimize model performance.

Results: Among the tested models, both the XGBoost and Random Forest models demonstrated superior performance across all metrics. The models achieved a precision of 0.87, recall of 0.89, and a F1 score of 0.88 for the positive class (n = 13,147 samples), aligning with findings reported in the literature. For the negative class (n = 9,840 samples), similar performance was observed with precision of 0.84, recall of 0.83, and a F1 score of 0.84.

Conclusions: Our study highlights the potential of machine learning in PV case causality classification, thereby supporting effective surveillance efforts. Future work should incorporate external domain expertise to validate findings and expand the dataset with additional features or examples to improve generalizability.