(A-181) Explainable Artificial Intelligence for Predicting Chemotherapy Toxicity: A Machine Learning Approach Using Pooled Clinical Trial Data

Background: Machine learning (ML) algorithms are increasingly used in health outcomes research for predictive modeling, yet their limited interpretability poses challenges for clinical decision-making. Explainable artificial intelligence (XAI) techniques can enhance understanding of ML predictions, improving transparency and trust in model-driven insights.

Objectives: To develop ML models for predicting 5-fluorouracil (5-FU)-based chemotherapy toxicity in metastatic colorectal cancer patients and to apply XAI techniques for model interpretation, thereby demonstrating their utility in health outcomes research.

Methods: A pooled analysis was conducted using individual patient data from five randomized clinical trials accessed through the Project Data Sphere. Patients with metastatic colorectal cancer who received 5-FU-based chemotherapy were included. Predictors considered for toxicity risk assessment included demographic variables, clinical history, and baseline laboratory values. Six predictive models were developed: k-nearest neighbors, naive Bayes, support vector machine, random forest, XGBoost, and logistic regression. Model performance was assessed using accuracy and the AUC through 10-fold cross-validation. For global interpretability, SHAP (SHapley Additive exPlanations) values, permutation feature importance, and partial dependence plots were employed to evaluate the impact of individual predictors on toxicity risk. Local interpretability was assessed using SHAP dependence plots, LIME, and ceteris-paribus profiles to examine individual predictions.

Results: XGBoost demonstrated the highest predictive performance (AUC = 0.86, accuracy = 80%), outperforming random forest (AUC = 0.82, accuracy = 79%) and the traditional logistic regression model (AUC = 0.73, accuracy = 72%). SHAP analysis revealed that age ≥70 years had the highest mean absolute SHAP value (0.24), followed by BMI >30 (0.22), cardiac comorbidities (0.17), and renal dysfunction (0.15), indicating their substantial influence on toxicity risk. Global XAI methods confirmed age and BMI as key drivers of model predictions. Local XAI techniques highlighted patient-specific risk profiles, supporting individualized risk stratification.

Conclusions: XAI techniques enhance the interpretability of ML models in predicting chemotherapy toxicity, improving their applicability in pharmacoepidemiology. These findings suggest that integrating XAI into predictive modeling may aid in personalized treatment planning and clinical decision support.