JMIR Publications

ABSTRACT

Background:

Machine learning models with time-series data, unlike one-snap data collection that only identifies high-risk patients, can predict the adverse event to aid cancer therapy's race against time.

Objective:

This study used a time-series data collecting method to develop and validate machine learning models for sunitinib- and sorafenib-associated thyroid dysfunction.

Methods:

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 (GBDT) were used to develop the models. Prediction performances were compared using the accuracy, precision, recall, f1 score, the area under the receiver operating characteristic curve (AUROC), and the area under the precision-recall curve (AUPRC). The optimal threshold of the best-performing model was selected based on the maximum f1 score. SHapley Additive exPlanations (SHAP) were applied to understand feature importance and contributions for the cohort and each patient.

Results:

The training cohort contained 609 patients, whereas the temporal validation cohort had 198 patients. The GBDT without resampling outperformed other models, with respective values of the AUPRC, AUROC, and f1 score of 0.600, 0.876, and 0.583 after adjusting the threshold. The SHAP analysis revealed that the most important features were a higher cholesterol level, longer summed days of medication use, and histology of clear cell adenocarcinoma. The final model was further integrated into a web-based application.

Conclusions:

The model can be an explainable adverse drug reaction surveillance system for predicting sunitinib- and sorafenib-associated thyroid dysfunction.