An Explainable Machine Learning Model with SHAP Feature Selection and Hybrid Sampling for Anemia Pre-Screening

The clinical differentiation process among suspected cases of hemoglobinopathy related anemia specifically Beta Thalassemia Trait (BTT), Hemoglobin E (HbE), Iron Deficiency Anemia (IDA), and their mixed forms is essential due to the differing clinical management and genetic counseling approaches required for each condition. Conventionally, this procedure involves manual examination of erythrocyte parameters for each suspected patient. However, when applied in resource-limited settings with a large patient population, this approach can become inefficient and time-consuming. To address this challenge, various studies have explored the use of Explainable Artificial Intelligence (XAI) technologies. Nevertheless, the application of such technologies in Indonesia remains limited, particularly in the context of primary patient data characterized by local demographic profiles. Moreover, the available clinical datasets often suffer from limitations, including small sample sizes and imbalanced class distributions, as is the case in this study.To overcome these challenges, a classification model was developed to enable pre-screening of anemia types based on Machine Learning with Explainable Artificial Intelligence (ML-XAI). The model integrates data-agnostic hybrid resampling techniques, including SMOTE-TOMEK and SMOTE-ENN, to address class imbalance, and employs SHAP-based feature importance for feature selection to enhance model performance. The model is designed to classify four types of anemia and to generate interpretable predictions that can be clinically analyzed by end users through XAI-based visualization.The model was trained on a dataset comprising 19 features derived from Complete Blood Count (CBC) tests conducted using Advia and Sysmex hematology analyzers. The resulting model demonstrated a recall performance of 88% for Beta Thalassemia Trait, 65% for Iron Deficiency Anemia, 100% for Hemoglobin E, and 43% for mixed anemia. Overall, the model achieved a macro recall of 74%, precision of 65%, F1-score of 67%, accuracy of 73%, and an AUC of 85%. All codes and data use in this research can be accessed in this URL http://ugm.id/AMLDSPermana