An Ensemble Method for the Analysis of Small Biomedical Data based on a Neural Network Without Training.

To enhance the accuracy of analyzing short datasets, this paper proposes a novel ensemble learning method that utilizes a single the General Regression Neural Network (GRNN). The core idea behind this method is the synthesis of additional pairs of vectors with different signs around each current vector from the test sample. This is achieved by employing the method of random symmetric perturbations and averaging the prediction outputs for the current vector and all synthesized vectors in its vicinity. Implementing this approach leads to a significant increase in prediction accuracy for short datasets. It achieves error compensation for each pair of additional vectors with different signs and also for the overall prediction result of the current vector and all additional pairs of synthetic vectors created for it. The effectiveness of the proposed method is validated through modeling on a small real-world biomedical dataset, and the optimal parameters have been selected. Comparative analysis with existing GRNN-based methods demonstrates a substantial improvement in accuracy. [ABSTRACT FROM AUTHOR]