Enhancing Breast Cancer Prediction with XAI-Enabled Boosting Algorithms
Vinayak Gupta, Richa Sharma · 2024
Breast cancer poses a significant health threat to the growing population of the world and as such calls for innovative methods to improve early detection and improve patient safety. The existing techniques in the field have made use of techniques like various image processing models like RetinaNet, You Only Look Once, Convolutional Neural Networks as well as tree-based models like the J48, decision trees and Naïve Bayes to predict Brest cancer but most these studies focus on the accuracy of the predictions without providing a basis for explanation and understandability of the results. Therefore, to provide transparency, this study utilizes the Local Interpretable Model-agnostic Explanations (LIME). This study investigated the performance of three boosting algorithms LightGBM, CatBoost and XGBoost in breast cancer prediction making use of the SEER dataset. The paper also makes use of SMOTE and the amalgamation of Tomek Links to handle class imbalance generally present with medical data to enhance the generalizability of the results. With hyperparameter tuning, the results show that LightGBM outperformed the other two in terms of accuracy as well as recall metrics. LIME explanations were then employed specific patient instances to understand the interplay of the various variables and how they contribute to the model predictions. These explanations highlighted the clinical relevance of variables such as T Stage and Grade of cancer, aiming to instill confidence among the clinicians. The research in all undermines the importance of integrating XAI methodologies in breast cancer detection, paving way for more clinically regulated and relevant models in the future.