Scholarlink

<p>Cardiovascular disease (CVD) prediction across heterogeneous clinical datasets remains challenging due to feature inconsistencies, missing values, and distribution shifts between cohorts. This study proposes a Hybrid Feature Tokenizer Transformer (HyFT) integrating a Feature Tokenizer Transformer with a residual Multilayer Perceptron branch for transferable CVD risk prediction. Framingham and Kaggle datasets were harmonized and merged for training, while the Cleveland dataset was strictly reserved for external validation to evaluate cross-cohort generalization. A comprehensive preprocessing pipeline including canonical feature mapping, imputation with missingness indicators, standardization, Salp Swarm Algorithm (SSA) based feature selection, Principal Component Analysis (PCA), and Particle Swarm Optimization (PSO) hyperparameter tuning was implemented. The proposed HyFT model achieved 0.93 accuracy and 0.92 Macro-F1 on the unseen Cleveland cohort, outperforming Support Vector Machine (SVM), Random Forest (RF), XGBoost, and Multi-Layer Perceptron (MLP) baselines. Ablation and bootstrap confidence interval analyses confirmed the statistical reliability and contribution of each component, demonstrating robust and clinically interpretable multi-cohort CVD prediction.</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

Hybrid FT-transformer with residual MLP for transferable cardiovascular risk prediction across cohorts

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月30日

查看全文
  • T. K. Revathi,
  • B. Sathiyabhama,
  • S. Kaliraj,
  • M. Dhasny Lydia,
  • V. Sivakumar

摘要

Cardiovascular disease (CVD) prediction across heterogeneous clinical datasets remains challenging due to feature inconsistencies, missing values, and distribution shifts between cohorts. This study proposes a Hybrid Feature Tokenizer Transformer (HyFT) integrating a Feature Tokenizer Transformer with a residual Multilayer Perceptron branch for transferable CVD risk prediction. Framingham and Kaggle datasets were harmonized and merged for training, while the Cleveland dataset was strictly reserved for external validation to evaluate cross-cohort generalization. A comprehensive preprocessing pipeline including canonical feature mapping, imputation with missingness indicators, standardization, Salp Swarm Algorithm (SSA) based feature selection, Principal Component Analysis (PCA), and Particle Swarm Optimization (PSO) hyperparameter tuning was implemented. The proposed HyFT model achieved 0.93 accuracy and 0.92 Macro-F1 on the unseen Cleveland cohort, outperforming Support Vector Machine (SVM), Random Forest (RF), XGBoost, and Multi-Layer Perceptron (MLP) baselines. Ablation and bootstrap confidence interval analyses confirmed the statistical reliability and contribution of each component, demonstrating robust and clinically interpretable multi-cohort CVD prediction.