An Explainable AI-Guided Framework for Cancer Pain Phenotyping and Clinical Decision-Making.

PURPOSE: Current pain management often relies primarily on pain intensity, which may not capture patient variability or guide individualized care. Advances in machine learning (ML) and explainable artificial intelligence (XAI) offer opportunities to improve assessment by identifying clinically meaningful subgroups and enhancing model interpretability. This study aimed to identify distinct cancer pain phenotypes using explainable machine learning and to develop a phenotype-guided clinical decision framework to support individualized pain assessment and management.

DESIGN

A cross-sectional study.

METHODS

Data from 260 cancer patients was analyzed using ML to identify distinct pain phenotypes. Variables included pain intensity, catastrophizing, psychological distress, quality of life, and self-management strategies. Three ML models, Extreme Gradient Boosting (XGB), Light Gradient Boosting Machine (LGBM), and Categorical Boosting (CatBoost), were evaluated using accuracy, precision, recall, and F1 score. Two XAI methods, SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME), were applied to enhance model transparency and identify key predictors of phenotype classification.

RESULTS

Four cancer pain phenotypes were identified: High-Burden Pain, Catastrophizing-Dominant, Stress-Related Pain, and Low-Risk Adaptive. CatBoost achieved the highest performance (accuracy = 0.98; F1 score = 0.97). SHAP analysis indicated that pain catastrophizing and psychological distress were the strongest predictors, whereas pain intensity contributed less. LIME analyses confirmed that individual predictions were primarily driven by cognitive and emotional factors. A phenotype-guided clinical decision framework was developed to align patient profiles with targeted interventions.

CONCLUSION

Integrating ML and XAI into cancer pain assessment provides clinically interpretable insights into multidimensional patient profiles. The identified phenotypes and proposed framework support personalized, mechanism-based pain management. Incorporating psychological and behavioral dimensions into routine assessment may improve clinical decision-making and patient outcomes.

CLINICAL IMPLICATION: Identifying distinct pain phenotypes may support individualized pain assessment and management by helping healthcare professionals recognize patient subgroups with different pain characteristics and treatment needs.