Scholarlink

<p>Pulmonary Hypertension (PH) is a critical cardiopulmonary disorder that is characterized by significant diagnostic delays. Artificial Intelligence (AI) offers promising non-invasive tools for early detection. This systematic review synthesizes AI/ML applications for PH detection, stratifies performance by World Health Organization (WHO) groups and data modalities, and critically evaluates the field’s progress. A unique aim is to assess the adoption of privacy-preserving collaborative techniques, particularly federated learning (FL), as a solution for the data fragmentation that impedes robust model development. A dual-stream review was conducted. First, a PRISMA 2020-compliant systematic literature search (2015–2025) identified studies applying machine learning (ML) and deep learning (DL) to PH detection across echocardiography, computed tomography, electrocardiography, and chest radiography; 50 studies were included. Study-level methodological quality was assessed using a modified QUADAS-2 framework. Second, a targeted benchmarking review examined federated learning (FL) adoption in analogous medical domains. Analysis of 50 PH-AI studies reveals that models, particularly DL applied to echocardiography and CT, attain high diagnostic accuracy (AUC 0.83–0.99). A pronounced research imbalance exists, with most studies focused on Group 1 pulmonary arterial hypertension (PAH) despite Group 2 (PH-LHD) having higher prevalence. Common limitations include small, single-center datasets and poor generalizability. A critical finding is the complete absence of FL in PH research, despite its proven success in overcoming identical data challenges in other fields. This FL gap represents a pivotal, addressable barrier. AI/ML shows significant promise for non-invasive PH detection but is critically hindered by data silos and a lack of generalizable studies. FL is identified as a promising but as-yet unimplemented translational direction to enable privacy-preserving multicenter collaboration. Future work must prioritize FL frameworks alongside the development of robust, multi-modal models that address all WHO PH groups to enable successful clinical translation.</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 Systematic Review of AI for Pulmonary Hypertension Detection: Performance, Gaps, and the Critical Need for Federated Learning

  • 期刊论文

  • 发表日期：

    2026年6月2日

查看全文
  • Angsar Taigonyrov,
  • Hari Mohan Rai,
  • Prashant Jamwal,
  • Aditya Pal,
  • Abdul Razaque

摘要

Pulmonary Hypertension (PH) is a critical cardiopulmonary disorder that is characterized by significant diagnostic delays. Artificial Intelligence (AI) offers promising non-invasive tools for early detection. This systematic review synthesizes AI/ML applications for PH detection, stratifies performance by World Health Organization (WHO) groups and data modalities, and critically evaluates the field’s progress. A unique aim is to assess the adoption of privacy-preserving collaborative techniques, particularly federated learning (FL), as a solution for the data fragmentation that impedes robust model development. A dual-stream review was conducted. First, a PRISMA 2020-compliant systematic literature search (2015–2025) identified studies applying machine learning (ML) and deep learning (DL) to PH detection across echocardiography, computed tomography, electrocardiography, and chest radiography; 50 studies were included. Study-level methodological quality was assessed using a modified QUADAS-2 framework. Second, a targeted benchmarking review examined federated learning (FL) adoption in analogous medical domains. Analysis of 50 PH-AI studies reveals that models, particularly DL applied to echocardiography and CT, attain high diagnostic accuracy (AUC 0.83–0.99). A pronounced research imbalance exists, with most studies focused on Group 1 pulmonary arterial hypertension (PAH) despite Group 2 (PH-LHD) having higher prevalence. Common limitations include small, single-center datasets and poor generalizability. A critical finding is the complete absence of FL in PH research, despite its proven success in overcoming identical data challenges in other fields. This FL gap represents a pivotal, addressable barrier. AI/ML shows significant promise for non-invasive PH detection but is critically hindered by data silos and a lack of generalizable studies. FL is identified as a promising but as-yet unimplemented translational direction to enable privacy-preserving multicenter collaboration. Future work must prioritize FL frameworks alongside the development of robust, multi-modal models that address all WHO PH groups to enable successful clinical translation.