Scholarlink

<p>Accurately classifying cancer subtypes from high-dimensional gene expression data is still difficult due to noise, redundancy, and the absence of comprehensive frameworks. The goal of this work is to create an understandable and computationally effective pipeline that enhances classification performance across various cancer datasets while ensuring generalizability and robustness. To find stable gene subsets, the suggested approach combines hybrid feature selection with L1-Logistic Regression, Variance-ANOVA selector, Random Forest importance, and Mutual Information. Fuzzification of selected genes is followed by multi-threshold FP-Growth for rule extraction. A Bi-Branch Attention DNN combined with Logistic Regression fuses discretized rule patterns with fuzzified features. To assess cross-dataset consistency, experiments are carried out on five datasets, including both microarray and proteomic modalities. Comparing the hybrid technique to classic classifiers, it improves predictive performance on all datasets. The Bi-Branch Attention DNN with Logistic Regression reliably achieves the highest accuracy after feature selection and rule-guided representation. The model’s short inference speeds across datasets, and a lightweight architecture facilitates practical applicability. Statistical feature selection, fuzzy rule mining, and dual-branch deep learning are used to create a successful cancer subtype prediction technique. The suggested method achieves strong performance and good generalization across omics types while preserving computational efficiency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Cancer Subtype Classification Using Hybrid Feature Selection and Bi-Branch Attention Deep Learning

  • 期刊论文

  • 发表日期：

    2026年5月13日

查看全文
  • Maisha Fahmida,
  • Shawly Ahsan,
  • Mohammad Shamsul Arefin

摘要

Accurately classifying cancer subtypes from high-dimensional gene expression data is still difficult due to noise, redundancy, and the absence of comprehensive frameworks. The goal of this work is to create an understandable and computationally effective pipeline that enhances classification performance across various cancer datasets while ensuring generalizability and robustness. To find stable gene subsets, the suggested approach combines hybrid feature selection with L1-Logistic Regression, Variance-ANOVA selector, Random Forest importance, and Mutual Information. Fuzzification of selected genes is followed by multi-threshold FP-Growth for rule extraction. A Bi-Branch Attention DNN combined with Logistic Regression fuses discretized rule patterns with fuzzified features. To assess cross-dataset consistency, experiments are carried out on five datasets, including both microarray and proteomic modalities. Comparing the hybrid technique to classic classifiers, it improves predictive performance on all datasets. The Bi-Branch Attention DNN with Logistic Regression reliably achieves the highest accuracy after feature selection and rule-guided representation. The model’s short inference speeds across datasets, and a lightweight architecture facilitates practical applicability. Statistical feature selection, fuzzy rule mining, and dual-branch deep learning are used to create a successful cancer subtype prediction technique. The suggested method achieves strong performance and good generalization across omics types while preserving computational efficiency.