Anemia diagnosis from clinical records leveraging tree-driven and XAI-enhanced machine learning

South Asian countries are highly vulnerable to anemia- a condition affecting numerous children, women, and adolescents. Early detection and effective treatment of anemia could significantly reduce its prevalence. Therefore, in this study, we leveraged clinical records from a Bangladeshi Hospital to develop an adaptive machine learning framework designed to predict anemia by analyzing patients' hematological parameters. Our research includes comprehensive data exploration and introduces XAI-enhanced Tree-based models, which comprises four traditional machine learning algorithms: Decision Tree Classifier, Gradient Boosting Decision Trees, Random Forest Classifier, and Extra Trees Classifier. We addressed key data analytics challenges using various statistical methods, including outlier management with the Z-score test, feature skewness (and kurtosis) analysis, feature correlation assessment via the Pearson correlation technique, feature importance evaluation with the Random Forest Classifier, and a Z-test to determine the p-value of anemia-related features. Besides, the machine learning models are optimized by applying Grid Search hyperparameter tuning. Our fine-tuned Tree-driven models, particularly Random Forest Classifier and Gradient Boosting Decision Trees, demonstrated exceptional predictive performance in all the evaluation metrics. We also performed stratified cross-validation, computed 95 % bootstrap confidence intervals, and conducted permutation-based validation, which yielded a minimal p-value-all to ensure a rigorous evaluation of our models. Furthermore, exploratory data analysis and SHAP explainable plots revealed that 'hemoglobin level' is the most important risk factor with substantial influence on anemia diagnosis and decision-making. We utilized a parallel open-source dataset to validate our model's performance, and the experimental results highlighted our proposed pipelines superiority. Moreover, to support adoption in clinical settings, we developed an intuitive real-time prediction interface that provides a clinical risk score along with SHAP-based model explanations, enabling healthcare professionals to understand the reasoning behind each prediction. To conclude, we believe the introduced framework could serve as an AI-driven decision support system for anemia diagnosis, potentially offering a significant contribution to automating the healthcare industry.