Precision-calibrated LightGBM machine learning model to predict serious adverse events in oncology patients using FAERS.

Abstract

12017 Background: The U.S. Food and Drug Administration’s Adverse Event Reporting System (FAERS) is a large national repository capturing postmarketing drug safety events. Predicting serious adverse events (SAEs) is especially critical for oncology patients, who often face complex regimens and higher toxicity risks. Currently, there are no well-established predictive tools to identify those at highest risk of SAEs in the cancer population. We aimed to develop a machine learning (ML) model that leverages FAERS data to address this clinical gap. Methods: We collated FAERS records from 2012Q4–2024Q3 that listed cancer as an indication, encompassing demographics, outcome codes (SAEs defined as death, life-threatening, disability, hospitalization, or congenital anomaly/birth defect), and drug data. Excluding non-cancer or incomplete entries, we split the dataset into 80% for training and 20% for testing, applying cross-validation to derive 95% confidence intervals. Training data underwent SMOTETomek oversampling to counter class imbalance. We then built a precision-focused LightGBM model using high-depth RandomizedSearchCV to fine tune the hyperparameters and subsequently applied sigmoid calibration to yield stable probability estimates and enhanced interpretability. Additionally, a logistic regression model was built under the same pipeline as a baseline comparator. Results: Of the final 2.28 million oncology-related reports, ~44% met SAE criteria. On the ~450,000-report test subset, our calibrated LightGBM model achieved 75% accuracy, with precision = 73.7% and recall = 86.3%, yielding an F1 of 0.795. The AUROC reached ~0.82 (95% CI ~0.80–0.84), underscoring robust discrimination, and the AUPRC approached 0.77. By comparison, logistic regression attained 73% accuracy (precision 72.9%, recall 81.5%, F1 = 0.77). Thus, the LightGBM pipeline offered notable gains in recall and F1 while preserving practical precision. Cross-validation demonstrated stable performance (±~1–2% across folds), and preliminary partial SHAP analysis indicated that older age (≥65 years), multi-agent chemotherapy, and prior adverse event histories were among the strongest predictors of SAE risk. Conclusions: In this largest FAERS-based oncology SAE analysis, our LightGBM model markedly outperformed logistic regression, achieving high recall (86%) and balanced precision (74%). These findings represent a significant advance in real-world pharmacovigilance, enabling earlier and more reliable detection and prediction of severe toxicities in cancer patients. Future prospective validations, potentially incorporating external datasets, may further amplify its clinical impact.