UCLA Library

Predicting drug side effects is crucial for drug safety and development. In our project, we extend a prior study on ML-based drug side effect prediction by (1) evaluating additional domain knowledge features (higher resolution ATC codes) and (2) optimizing the logistic regression (LR) model. After replicating the original study, we assess ATC levels 3-5 and fine-tune LR model parameters. Results show levels 2-4 perform comparably, while level 5 underperforms, likely due to overfitting. Adjusting decision thresholds and regularization also improved performance compared to the original study. Our study highlights the tradeoff between feature specificity and model generalizability, emphasizing the need for careful feature selection and parameter tuning. These insights can further ML-based drug safety research, informing future pharmacological studies that integrate computational approaches for improved side effect prediction.