Machine Learning Model for Predicting Severe Adverse Events in Oncology Patients Using the US Food and Drug...

Area of Science:

• Pharmacovigilance and Machine Learning
• Oncology Drug Safety
• Real-World Data Analysis

Background:

• Predicting severe adverse events (SAEs) in oncology is complex due to varied treatments and patient factors.
• Traditional pharmacovigilance methods struggle to identify multifactorial risk patterns for SAEs.
• Machine learning (ML) offers a promising approach to detect subtle SAE predictors in large datasets like FAERS.

Purpose of the Study:

• Develop and validate an ML model to predict severe oncology-related adverse events.
• Identify key risk factors contributing to the severity of adverse events in cancer patients.
• Utilize the US Food and Drug Administration Adverse Event Reporting System (FAERS) for comprehensive analysis.

Main Methods:

• Analyzed over 3.7 million oncology-related FAERS cases (2012Q4-2024Q3) with extensive data preprocessing.
• Defined severe events by outcomes: death, hospitalization, disability, congenital anomaly, or life-threatening conditions.
• Trained a LightGBM model, optimized with Optuna, and benchmarked against logistic regression, using SHAP for interpretability.

Main Results:

• The LightGBM model significantly outperformed logistic regression in predicting SAEs (AUROC 0.806 vs. 0.708).
• Key predictors identified include advanced age, polypharmacy (≥15 drugs), longer therapy duration, and higher number of reported reactions.
• SHAP analysis confirmed that age, polypharmacy, and therapy duration synergistically elevate SAE risk.

Conclusions:

• The gradient boosting model enhances prediction and interpretability of severe oncology adverse events.
• Clinically meaningful predictors identified via SHAP enable precision pharmacovigilance and targeted risk mitigation.
• Integration of ML into regulatory and clinical workflows can improve postmarket safety surveillance for oncology drugs.