Abstract
Hospital readmission within 30 days remains a significant issue in oncology practice, resulting in increased healthcare costs and adverse patient outcomes. Existing predictive models for breast cancer readmission often suffer from limited interpretability and generalisability. This preliminary study investigates the feasibility of developing and evaluating explainable machine learning (XAI) models to enhance the transparency of 30-day readmission prediction in patients with breast cancer.
Due to data availability constraints in the breast cancer readmission, a small-scale retrospective dataset of 100 patient records was analysed. The dataset comprises demographic, clinical, and treatment-related variables, including age, comorbidities, Eastern Cooperative Oncology Group (ECOG) performance status, baseline neutrophil count, and dosage adjustments. After conducting an exploratory analysis of the distribution of variables, several machine learning algorithms, including Logistic Regression, Support Vector Machines, Naïve Bayes, K-Nearest Neighbours, Decision Trees, Random Forest, and XGBoost, were implemented and evaluated for their performance. The ML algorithms were trained both with and without the use of the Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance. The data split ratio was 80% for training and 20% for testing, resulting in a limited test set of 20 samples.
Models achieved a moderate performance with the Random Forest (RF), with an accuracy of 75% and an F1-score of 83%. While these results suggest promising predictive potential within the study cohort, they are interpreted as preliminary evidence rather than definitive clinical metrics, given the sample size. Among the evaluated models, Naïve Bayes (NB) and Decision Tree (DT) also performed reasonably well, achieving an accuracy of 65% each. In contrast, Logistic Regression (LR), Support Vector Machine (SVM), and K-Nearest Neighbours (KNN) showed slightly lower performance. The findings suggest that tree-based models were superior to conventional statistical methods. The highest-performing balanced SMOTE Random Forest model achieved moderate accuracy, recall, and specificity, making it the most dependable classifier.
Interpretation analysis using Shapley Additive Explanations (SHAP), Partial Dependence Plots (PDP), and Local Interpretable Model-agnostic Explanations (LIME) indicated that comorbidity burden and ECOG and comorbidities performance status consistently emerged as the primary predictors of readmission. These findings align with established clinical knowledge, supporting the methodological validity of the models. From a practical perspective,...