Exploring the similarities of XAI approaches in finding the influential factors in lung cancer survival rate

Abstract

Lung cancer is the most common cancer and the leading cause of cancer-related deaths globally. While Artificial Intelligence (AI) offers promise in early detection of high-risk patients and providing treatment decision support, the black-box nature of AI models raises trust concerns. Therefore, eXplainable Artificial Intelligence (XAI) is important to explain models and identify critical factors affecting lung cancer survival. In addition, real datasets frequently contain missing values due to diverse reasons. In this work, the Simulacrum dataset was used, and different statistical and machine learning imputation methods like Mean-Mode, K-Nearest Neighbors (KNN), MissForest were used to impute the missing data. An innovative neural network imputation method was proposed to evaluate whether the sequence of imputing the features with missing values matters. Neural network models were then trained on the imputed datasets to predict lung cancer patient survival. The results revealed that the models trained on imputed datasets obtained from machine learning imputation methods had better performance and the sequence in which features were imputed had minimal impact on the performance. For interpretability, different XAI approaches were deployed, namely, Expected Gradients (EG), SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME). The analysis revealed that important features were related to the tumor condition, age and dose administration, largely unaffected by different imputation methods. Pairwise similarity metrics such as Jaccard Similarity, Pearson Correlation Coefficient and Cosine Similarity were used to compare the XAI approaches. EG and SHAP exhibited the highest similarity to each other, though there were some notable inconsistencies between their explanations. This paper underscores the usefulness of XAI for medical professionals in understanding model decisions and emphasizes the importance of cross-validating explanations from various XAI approaches to evaluate model reliability.