<![CDATA[The early detection of hematological disorders remains challenging because many conditions share similar clinical characteristics and show substantial variation in laboratory measurements. Existing machine learning systems often struggle to maintain consistent accuracy in multi-class settings with imbalanced data. The research contribution is a multi-class diagnostic framework that identifies nine hematological disease categories using only routine laboratory parameters, supported by a leakage-free evaluation protocol and a comprehensive comparison across baseline classifiers. The proposed solution uses an extreme gradient boosting model as the primary classifier and evaluates it against support vector machine, random forest, and extra trees. The method includes data cleaning and numerical standardization, and class balancing with the Synthetic Minority Oversampling Technique applied only to the training subset within each fold of ten-fold cross-validation to prevent optimistic bias. Model performance is assessed using accuracy, precision, recall, and F1-score, together with computational efficiency measured through processing time and memory usage. The results show that the extreme gradient boosting model achieves the best overall performance, with an average accuracy of 98.67%, precision of 98.80%, recall of 98.67%, and an F1-score of 98.66%. It also demonstrates efficient memory usage and shorter processing time compared with the other tested methods. The competing models perform adequately but exhibit higher variability and weaker recognition for minority classes. In conclusion, these findings indicate that extreme gradient boosting provides an accurate and efficient approach for hematology-based multi-class disease classification when evaluated under a strict, leakage-free resampling protocol.]]>
