Parameter optimization of machine learning algorithms using metaheuristic methods in classifying anemia disease

Anemia occurs when the hemoglobin (Hgb) level falls below a certain reference range. Diagnosis and treatment processes require many blood tests, radiological imaging and various tests. Using machine learning methods, patients' medical data can be processed to make disease predictions for new patients and provide decision support mechanisms to doctors based on these predictions. These methods are critical in reducing the margin of error in doctors' diagnoses, and the evaluation of data records in healthcare institutions is of great importance for both patients and hospitals. This thesis aims to evaluate the performance of various models on both two-class (anemia present/absent) and multi-class (non-patient, HGB-anemia, iron deficiency anemia, B12 deficiency anemia, folate deficiency anemia) anemia datasets. For this purpose, data imbalance in the datasets was eliminated with the SMOTE (Synthetic Minority Oversampling Technique) method, then tests were performed with classical methods frequently used in the literature, and the datasets were linearized to better model the relationship between the parameters in the datasets and to emphasize the interaction of the parameters with each other, The parameters were modeled in quadratic and exponential form and parameter optimization was performed with the Harris Hawk Algorithm (HSA), Crow Search Algorithm (KAA), Chicken Swarm Optimization Algorithm (TSO), JAYA Algorithm (JAYA), Whale Optimization Algorithm (BOA) and Butterfly Optimization Algorithm (KOA) metaheuristics. Emphasizing the need to increase the class-based accuracy as well as the overall accuracy in a multi-class data set, hybrid models are proposed by combining the classical method TreeBagger (TB), which better emphasizes the importance of parameters inspired by the fuzzy logic approach, and the proposed metaheuristic methods.