Web-Based Explainable Machine Learning-Based Drug Surveillance for Predicting Sunitinib- and Sorafenib-Associated...

Unlike one-snap data collection methods that only identify high-risk patients, machine learning models using time-series data can predict adverse events and aid in the timely management of cancer.

This study aimed to develop and validate machine learning models for sunitinib- and sorafenib-associated thyroid dysfunction using a time-series data collection approach.

Time series data of patients first prescribed sunitinib or sorafenib were collected from a deidentified clinical research database. Logistic regression, random forest, adaptive Boosting, Light Gradient-Boosting Machine, and Gradient Boosting Decision Tree were used to develop the models. Prediction performances were compared using the accuracy, precision, recall, F1-score, area under the receiver operating characteristic curve, and area under the precision-recall curve. The optimal threshold for the best-performing model was selected based on the maximum F1-score. SHapley Additive exPlanations analysis was conducted to assess feature importance and contributions at both the cohort and patient levels.

The training cohort included 609 patients, while the temporal validation cohort had 198 patients. The Gradient Boosting Decision Tree model without resampling outperformed other models, with area under the precision-recall curve of 0.600, area under the receiver operating characteristic curve of 0.876, and F1-score of 0.583 after adjusting the threshold. The SHapley Additive exPlanations analysis identified higher cholesterol levels, longer summed days of medication use, and clear cell adenocarcinoma histology as the most important features. The final model was further integrated into a web-based application.

This model can serve as an explainable adverse drug reaction surveillance system for predicting sunitinib- and sorafenib-associated thyroid dysfunction.