<![CDATA[Complete Blood Count (CBC) remains the cornerstone for initial screening of hematological disorders, yet manual interpretation is often challenged by overlapping biological patterns and substantial inter-patient variability. Although machine learning approaches have demonstrated promise for automated diagnosis, many existing studies prioritize classification accuracy while neglecting computational efficiency and the persistent class imbalance inherent in medical datasets. This study develops a lightweight yet effective diagnostic framework for classifying nine hematological conditions using routine CBC parameters. Evaluated on a public dataset of 1,281 records from Kaggle, the proposed model is benchmarked against standard Random Forest, XGBoost, and Support Vector Machine (SVM) classifiers. The approach integrates the Synthetic Minority Oversampling Technique (SMOTE) to mitigate class imbalance, and Binary Particle Swarm Optimization (BPSO) to identify a compact and clinically informative feature subset of exactly 6 parameters, referred to as a clinical fingerprint, optimized for the Extra Trees classifier. Evaluated using ten-fold cross-validation, the BPSO-Extra Trees model achieved an average accuracy of 87.43 percent and an F1 score of 82.75 percent, while demonstrating superior resource efficiency with peak memory consumption of only 0.249 MB, corresponding to a 46.6 percent reduction compared with the standard Random Forest baseline. These findings confirm that swarm intelligence optimized models can effectively balance diagnostic performance with extreme computational frugality, enabling the potential deployment of accurate hematology-based decision support systems on portable devices and in resource-limited laboratory environments.]]>
