Scholarlink

<p>Intradialytic hypotension (IDH) prediction requires modeling static clinical context and long dialysis time-series under class imbalance and tight latency constraints. We propose CNN–Mamba–WOA, a dual-branch framework that combines a CNN encoder for static variables, a Mamba state-space branch with linear-time sequence complexity, and a multi-objective Whale Optimization Algorithm that tunes hyperparameters by jointly optimizing accuracy, ROC–AUC, F1, MCC, and G-Mean. Implemented with GPU-accelerated mini-batch training, the framework naturally maps to multi-GPU and cluster environments. Experiments on a public cohort of 758 patients and 98,015 sessions show that CNN–Mamba–WOA achieves 0.823 ROC–AUC and 0.902 accuracy, improving ROC–AUC by 0.044 over the best CNN–LSTM baseline. SHAP-based feature attributions and temporal saliency maps further provide clinically meaningful explanations, enabling near real-time, supercomputing-ready decision support for hemodialysis units.</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

CNN–Mamba–WOA: an efficient and explainable state-space fusion framework with multi-objective optimization for large-scale hemodialysis time-series prediction

  • 期刊论文

  • 发表日期：

    2025年12月24日

查看全文
  • Weihua Pan,
  • Yujie Liu

摘要

Intradialytic hypotension (IDH) prediction requires modeling static clinical context and long dialysis time-series under class imbalance and tight latency constraints. We propose CNN–Mamba–WOA, a dual-branch framework that combines a CNN encoder for static variables, a Mamba state-space branch with linear-time sequence complexity, and a multi-objective Whale Optimization Algorithm that tunes hyperparameters by jointly optimizing accuracy, ROC–AUC, F1, MCC, and G-Mean. Implemented with GPU-accelerated mini-batch training, the framework naturally maps to multi-GPU and cluster environments. Experiments on a public cohort of 758 patients and 98,015 sessions show that CNN–Mamba–WOA achieves 0.823 ROC–AUC and 0.902 accuracy, improving ROC–AUC by 0.044 over the best CNN–LSTM baseline. SHAP-based feature attributions and temporal saliency maps further provide clinically meaningful explanations, enabling near real-time, supercomputing-ready decision support for hemodialysis units.