A dynamic rescaled activation kernel network for chaotic pattern recognition and early disability risk mitigation as a biomarker in cancer classification

Abstract

Activation functions and their weight adjustment in back propagation play a crucial role in digital decision systems, particularly in early diagnosis of diseases and capturing chaotic patterns prevalent in medical datasets. Conventional activation functions pose significant challenges that hinder the performance and scalability of machine learning models. In this research, Binary-Scaled Sigmoid Activation Function (BSSAF) based on the exponential family is proposed to address the bifurcation Instability and sensitivity at outliers to reduce the loss of decisive information in the complex pattern. The rescaled activation function enhances gradient flow during backpropagation by converting asymptotic noise into a robust signal, effectively restoring the information process. BSSAF in the machine learning paradigm extracts the hidden information more precisely and captures the weak decision boundaries. Special transformations are also introduced to reshape non-Gaussian patterns into a normal distribution to enhance symmetry and convergence of the gradient-based optimization algorithm. The performance of the binary classifier is evaluated for the imbalanced medical dataset. Cancer is a leading disease worldwide, affecting the health of millions of people each year. The BSSAF, with its proposed transformation, is applied to diagnose breast cancer to avoid long-term disability. Based on enhanced performance, the design paradigm can help develop screening tools to reduce mortality by improving the precision of the health information system. Comparative performance analysis for different classification algorithms, including Logistic, SVM, and Xgboost, is presented to evaluate the accuracy based on the breast cancer dataset. The experimental result confirmed that the BSSAF possesses superior performance as compared to other activation functions, with an F1 score of 99 %. The dynamic bifurcation ability of the binary scaled activation function can be utilized further for medical images, time series pattern identification to achieve high accuracy and precision.