Optimization of dendritic neuron model framework through activation pairing for multidomain medical diagnosis

Accurate and interpretable medical diagnosis remains a critical challenge due to the heterogeneous nature of clinical data and the limitations of conventional machine learning models in capturing complex nonlinear relationships. Dendritic neuron models (DNMs), inspired by biological neural processing, offer a promising alternative through localized nonlinear integration. Rather than introducing new dendritic architectures, this study presents a systematic and activation-aware analysis of existing dendritic neuron models to examine how activation function pairings and dendritic depth influence learning stability and classification performance. Three dendritic variants, namely the Standard DNM, Multi-Dendritic Neural Network (MDNN), and Multi-In and Multi-Out Dendritic Neuron Layer (MODN), are evaluated under multiple dendritic–somatic activation pairings using a unified gradient-based learning framework. Unlike prior studies that rely on metaheuristic optimization or task-specific tuning, gradient-based training is adopted to improve convergence efficiency and reproducibility. Experimental results on five heterogeneous medical datasets demonstrate consistently strong diagnostic performance, achieving accuracy above 90 % on anemia and breast cancer datasets and competitive results on heart disease, diabetes, and hepatitis. Overall, the study provides insight into the interaction between activation behavior and dendritic architecture, highlighting the importance of activation-aware modeling for biologically inspired medical diagnosis and establishing a foundation for efficient and deployable dendritic learning systems.