Scholarlink

<p>This systematic review explores the contribution of transfer learning to the prediction of cardiovascular diseases, highlighting deep neural network frameworks and the integration of multimodal data. In accordance with the PRISMA 2020 framework, a systematic search was performed in PubMed, IEEE Xplore, and ScienceDirect from 2017 to 2025, resulting in seven studies that fulfilled the inclusion requirements (≈ 6000 patients). Retrieved parameters comprised model type, dataset features, sample size (n), and performance indicators like accuracy, sensitivity, specificity, and AUROC (95%). Among the methods reported, the DANomaly Generative Adversarial Convolution Neural Network (DGACNN) model reached 85% accuracy for classifying congenital heart disease, while a dual-input ECG + PCG network achieved 94.7% accuracy (AUROC 0.95 [0.92–0.97]). Image-driven models like VGG-MI2 showed the best diagnostic effectiveness, achieving 99.2% accuracy (AUROC 0.99 [0.98–1.00]) in identifying myocardial infarction. Sample size, modality, and validation strategy impacted performance variation. Numerous studies did not have clear controls for patient-level data leakage, signal preprocessing, and model calibration, which could inflate the reported metrics. In spite of these constraints, transfer learning reliably improved diagnostic precision and generalization, especially when utilizing pretrained backbones and combining multiple modalities. Subsequent studies need to include external validation, reporting based on threshold Sn/Sp and PPV/NPV, along with decision-curve or calibration analyses to guarantee clinical relevance. Implementing standardized assessment according to Prediction Model Risk of Bias Assessment Tool (PROBAST) and Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis – Artificial Intelligence Extension (TRIPOD-AI) guidelines is crucial for converting these models into dependable, practical cardiovascular screening instruments.</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 comprehensive review of transfer learning techniques for predicting heart diseases

  • 期刊论文

  • 发表日期：

    2026年2月13日

查看全文
  • Farhan Masud Nayem,
  • Farida Siddiqi Prity

摘要

This systematic review explores the contribution of transfer learning to the prediction of cardiovascular diseases, highlighting deep neural network frameworks and the integration of multimodal data. In accordance with the PRISMA 2020 framework, a systematic search was performed in PubMed, IEEE Xplore, and ScienceDirect from 2017 to 2025, resulting in seven studies that fulfilled the inclusion requirements (≈ 6000 patients). Retrieved parameters comprised model type, dataset features, sample size (n), and performance indicators like accuracy, sensitivity, specificity, and AUROC (95%). Among the methods reported, the DANomaly Generative Adversarial Convolution Neural Network (DGACNN) model reached 85% accuracy for classifying congenital heart disease, while a dual-input ECG + PCG network achieved 94.7% accuracy (AUROC 0.95 [0.92–0.97]). Image-driven models like VGG-MI2 showed the best diagnostic effectiveness, achieving 99.2% accuracy (AUROC 0.99 [0.98–1.00]) in identifying myocardial infarction. Sample size, modality, and validation strategy impacted performance variation. Numerous studies did not have clear controls for patient-level data leakage, signal preprocessing, and model calibration, which could inflate the reported metrics. In spite of these constraints, transfer learning reliably improved diagnostic precision and generalization, especially when utilizing pretrained backbones and combining multiple modalities. Subsequent studies need to include external validation, reporting based on threshold Sn/Sp and PPV/NPV, along with decision-curve or calibration analyses to guarantee clinical relevance. Implementing standardized assessment according to Prediction Model Risk of Bias Assessment Tool (PROBAST) and Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis – Artificial Intelligence Extension (TRIPOD-AI) guidelines is crucial for converting these models into dependable, practical cardiovascular screening instruments.