Impact of activation functions on dendritic neuron models in medical datasets

This study evaluates the dendritic neuron model (DNM) as an innovative approach for analysing medical datasets with balanced and imbalanced class distributions, encompassing breast cancer Wisconsin, hepatitis, heart failure clinical records (HFCR), anemia, and heart disease. The objective is to optimize DNM generalization through combinations of sigmoid, tanh, ReLU, and ELU activation functions, with a focus on detecting minority classes critical for clinical applications. Experiments employ stratified K-fold cross-validation to ensure robust evaluation. The sigmoid-ReLU combination excels on imbalanced datasets, achieving an accuracy of 0.9807 ± 0.0166 on breast cancer Wisconsin and an F1-score of 0.9179 ± 0.0453 on hepatitis, demonstrating robust handling of class imbalance. On balanced datasets, ReLU-ReLU achieves a perfect accuracy of 1.0000 ± 0.0000 on anemia, while ReLU-sigmoid stands out on heart disease with a ROC-AUC of 0.9026 ± 0.0576, reflecting effective class discrimination. Conversely, ELU-tanh exhibits limitations on HFCR, with an accuracy of 0.7092 ± 0.0740, indicating challenges with extreme class imbalance. This research provides theoretical insights into activation function dynamics in DNM and practical contributions for clinical decision support systems, with broad potential for heterogeneous medical data analysis.