Scholarlink

<p>Machine learning has gained significant traction in medical applications, particularly for disease risk prediction. Its effectiveness, however, is frequently compromised by two interrelated challenges in clinical datasets, namely class imbalance and feature overlap. Class imbalance biases classifiers toward the majority-class, while overlap feature distributions blur inter-class decision boundaries. To address these dual constraints synergistically, this study proposes a novel hybrid framework integrating a triplet-enhanced autoencoder (TEA) with conditional generative adversarial networks (CGANs). The TEA module employs triplet loss constraints to learn discriminative feature representations, effectively mitigating feature overlap by enhancing inter-class margins in the latent space. These refined representations then condition the CGANs component, enabling the generation of high-quality minority-class samples that faithfully preserve pathological characteristics. Comprehensive evaluation across five diverse medical datasets demonstrates TEA-CGAN’s unique capacity to simultaneously resolve class imbalance and feature overlap.</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

A novel TEA-CGAN framework for mitigating class imbalance in disease risk prediction

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Ying Pan,
  • Bai Peng,
  • Mengying Xu,
  • Li Ma,
  • Limei Peng

摘要

Machine learning has gained significant traction in medical applications, particularly for disease risk prediction. Its effectiveness, however, is frequently compromised by two interrelated challenges in clinical datasets, namely class imbalance and feature overlap. Class imbalance biases classifiers toward the majority-class, while overlap feature distributions blur inter-class decision boundaries. To address these dual constraints synergistically, this study proposes a novel hybrid framework integrating a triplet-enhanced autoencoder (TEA) with conditional generative adversarial networks (CGANs). The TEA module employs triplet loss constraints to learn discriminative feature representations, effectively mitigating feature overlap by enhancing inter-class margins in the latent space. These refined representations then condition the CGANs component, enabling the generation of high-quality minority-class samples that faithfully preserve pathological characteristics. Comprehensive evaluation across five diverse medical datasets demonstrates TEA-CGAN’s unique capacity to simultaneously resolve class imbalance and feature overlap.