<![CDATA[Anemia remains a significant global health concern, and its diagnosis through manual interpretation of Complete Blood Count (CBC) results is susceptible to bias and misinterpretation. Machine learning techniques offer a promising solution for identifying complex patterns in medical data; however, their performance is often hindered by class imbalance issues commonly found in healthcare datasets. Therefore, this study aims to evaluate and compare the classification performance of Support Vector Machine (SVM) and Extreme Learning Machine (ELM) algorithms enhanced with the Synthetic Minority Over-sampling Technique (SMOTE) for anemia detection. The proposed approach employs SVM with an RBF kernel and ELM with a sigmoid activation function, both optimized through a 5-Fold Cross-Validation inner loop to determine the best regularization parameter C and the optimal number of hidden neurons, respectively. SMOTE is applied exclusively to the training data to address class imbalance without data leakage. To ensure robust and unbiased performance estimation, experiments were conducted over 10 independent runs using stratified 80:20 train-test splits on a secondary CBC dataset consisting of 200 patient records. Experimental results show that the SMOTE-based SVM model achieved a mean accuracy of 87.75% ± 0.75%, precision of 98.08% ± 1.92%, recall of 84.83% ± 1.69%, and F1-score of 90.94% ± 0.57%, with an average computation time of 1.430 seconds per run. In comparison, the SMOTE-based ELM model attained a mean accuracy of 85.25% ± 5.86%, precision of 95.81% ± 3.11%, recall of 83.45% ± 8.56%, and F1-score of 88.91% ± 4.89%, while requiring only 0.460 seconds per run. The findings suggest that SVM delivers more stable performance with higher precision, making it effective in minimizing false positive predictions. In contrast, ELM offers greater computational efficiency, making it suitable when rapid processing is prioritized over predictive accuracy.]]>
Copyrights © 2026
