Randomized Dendritic Neuron Model Preserving Biologically Inspired Dendritic Gating for Efficient Medical Diagnostics

Abstract

Non-communicable diseases (NCDs) such as type 2 diabetes, breast cancer, and cardiovascular diseases remain major global health burdens, requiring accurate yet computationally efficient predictive models. This study aims to develop a fast and scalable neuromorphic classifier for NCD prediction in resource-constrained clinical settings. We propose the Randomized Dendritic Neuron Model (R-DNM), a randomized variant of the Dendritic Neuron Model that preserves multiplicative dendritic interactions while fixing randomly initialized hidden-layer parameters and training only output weights via ridge regression. R-DNM was evaluated on three benchmark medical datasets using 10-fold cross-validation and compared against conventional DNM, its multi-output variant (MODN), and standard baselines including Random Forest, SVM, and MLP. On the Pima Indians Diabetes dataset, R-DNM achieved the best performance with an average accuracy of 0.7552 ± 0.0362, F1-score of 0.7529 ± 0.0343, and log-loss of 0.5668 ± 0.0218, outperforming DNM and MODN by up to 3.38% accuracy. Paired t-tests confirm significant improvements in F1-score (t = −3.77, p = 0.0014, Cohen’s d = 1.69) and computation time (p < 0.0001). These results establish R-DNM as a computationally efficient and statistically superior neuromorphic model for NCD prediction.