Scholarlink

Background <p>Women face unique cardiovascular disease risk stratification, mainly due to their body structure and hormone imbalances. Machine learning (ML)/solo deep learning (SDL) models are often ad-hoc and underperform. We hypothesize that hybrid deep learning (HDL) models designed by fusing unidirectional and bidirectional extended long-short term memory (xLSTM) embedded with gating are superior to SDL/ML.</p> Method <p>4 ML, 6 unidirectional SDL, and 12 bidirectional HDL models were designed in xLSTM framework. Feature engineering was conducted using differential expression analysis (DEA). Six types of scientific-validation paradigm were designed: (1) Unseen data analysis; (2) feature explainability; (3) memorization vs. generalization; (4) K-fold cross-validation; (5) reliability and stability analysis, and (6) benchmark HDL.</p> Results <p>(1) Mean percentage difference between seen and unseen analysis was 1.8% over 12 HDL models and 2.7% over six SDL models, respectively, meeting the regulatory requirements. The worst-case difference between seen and unseen for over 12 HDL and 6 SDL was 4.39% and 6.93%, respectively. On unseen data, the mean accuracy/AUC was 92.75% and 0.98, respectively over 12 HDL models, and 90.03% and 0.97, respectively over 6 SDL models. (2) On feature engineering, 80% features matched between Local Interpretable Model-agnostic Explanations (LIME) and DEA. (3) On generalization, HDL required 15% less data compared to SDL, (4) K-fold cross-validation showed consistent behaviour along all the models, (5) Reliability tests showed <i>p</i> value &lt; 0.01 for the model pairs. Compared to ML, SDL and HDL were superior by 10.17% and 12.31%.</p> Conclusions <p>Scientific validation and benchmarking demonstrated the reliability and robustness of proposed SDL and HDL models.</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

Scientific validation of hybrid deep learning using unseen paradigm and explainability for classification of women’s heart failure gene expression

  • 期刊论文

  • 发表日期：

    2026年3月18日

查看全文
  • Ekta Tiwari,
  • Dipti Shrimankar,
  • Krish Chaudhary,
  • Luca Saba,
  • Jasjit S. Suri

摘要

Background

Women face unique cardiovascular disease risk stratification, mainly due to their body structure and hormone imbalances. Machine learning (ML)/solo deep learning (SDL) models are often ad-hoc and underperform. We hypothesize that hybrid deep learning (HDL) models designed by fusing unidirectional and bidirectional extended long-short term memory (xLSTM) embedded with gating are superior to SDL/ML.

Method

4 ML, 6 unidirectional SDL, and 12 bidirectional HDL models were designed in xLSTM framework. Feature engineering was conducted using differential expression analysis (DEA). Six types of scientific-validation paradigm were designed: (1) Unseen data analysis; (2) feature explainability; (3) memorization vs. generalization; (4) K-fold cross-validation; (5) reliability and stability analysis, and (6) benchmark HDL.

Results

(1) Mean percentage difference between seen and unseen analysis was 1.8% over 12 HDL models and 2.7% over six SDL models, respectively, meeting the regulatory requirements. The worst-case difference between seen and unseen for over 12 HDL and 6 SDL was 4.39% and 6.93%, respectively. On unseen data, the mean accuracy/AUC was 92.75% and 0.98, respectively over 12 HDL models, and 90.03% and 0.97, respectively over 6 SDL models. (2) On feature engineering, 80% features matched between Local Interpretable Model-agnostic Explanations (LIME) and DEA. (3) On generalization, HDL required 15% less data compared to SDL, (4) K-fold cross-validation showed consistent behaviour along all the models, (5) Reliability tests showed p value < 0.01 for the model pairs. Compared to ML, SDL and HDL were superior by 10.17% and 12.31%.

Conclusions

Scientific validation and benchmarking demonstrated the reliability and robustness of proposed SDL and HDL models.