Abstract

In medical datasets class imbalance is an issue when the number of healthy cases significantly exceeds the number of Thalassemia cases and this hampers machine learning (ML) prediction accuracy. The focus of this paper is on the effectiveness of SMOTE and SHAP being combined in addressing imbalance and subsequently improving model interpretability for Thalassemia diagnosis through ML. The study involved training and testing five different algorithms, namely, SVM, Logistic Regression, Decision Tree, Random Forest, and XGBoost on an imbalanced set of patients. SMOTE was used to oversample the minority class thus balancing the dataset and reducing the predisposition towards the majority class. SHAP executed the pinpointing and dissecting of the most representative diagnostic features thereby the interpretability was improved dramatically. The evaluation revealed that SMOTE has remarkably enhanced the aspect of performance of the minority class; thus, using XGBoost the F1, score witnessed an increment of 60% (0.57 to 0.86) while AUC, ROC saw a rise of 21% (0.74 to 0.95). Altogether, the average F1, score and AUC, ROC were increased by approximately 41% and 21%, respectively. The upshot was SHAP brought out features corresponding to the clinical realm such as the concentration of hemoglobin and the total number of RBCs which is in line with medical knowledge. Physicians checked the SHAP output and confirmed their agreement with the diagnostic procedures. This approach of a combination from two different sides covers the challenge of imbalance in the data while at the same time enhancing interpretability thus augmenting the dependability and transparency of the model when used in the clinical perspective. The availability of our open, source code is an offer for the implementation of widespread thalassemia screening initiatives. 

Statistics