Scholarlink

<p>Federated learning (FL) is swiftly transforming the field of mental healthcare by allowing for multi-institutional, privacy-preserving analytics that do not require the centralization of sensitive patient data. In this systematic review, conducted in accordance with PRISMA guidelines, we synthesise evidence from 20 studies focused on FL for mental state detection (including emotion recognition), as well as 25-30 studies encompassing broader behavioural health and human activity applications published up to December 2024. Our work maps FL research across a comprehensive taxonomy—encompassing FL paradigm, data modality, mental health condition, aggregation strategy, scale, and real-world evaluation metrics, revealing the breadth and limitations of FL adoption in stress, depression, bipolar disorder, OCD, schizophrenia, emotion recognition, and other domains. FL models, such as FedAvg, CAFed, FedHealth, and FedUFO, achieve high diagnostic accuracy (85–95 %) on leading public datasets (e.g., WESAD, DAIC-WOZ), with only minor trade-offs (2–5 % decrease in accuracy) compared to centralized learning FL methods. Boosting model generalizability and robustness by exploiting data diversity across sites, and expanding integrations with blockchain technology reinforce auditability and trustworthiness. We provide further support to mental health professionals and policymakers on workflow integration, data governance, auditability, and multi-center collaboration, as well as a roadmap for practical deployment that includes user-centered evaluation and regulatory compliance. This review provides the first taxonomy-driven, methodologically rigorous overview of FL in mental health, critically evaluating the field’s current state, highlighting actionable paths for real-world adoption, and identifying open research opportunities that will guide the next generation of secure and inclusive mental health technologies.</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

Leveraging federated learning for mental healthcare: a systematic review

  • 期刊论文

  • 发表日期：

    2026年5月7日

查看全文
  • Lekha C Warrier,
  • Ragesh G K

摘要

Federated learning (FL) is swiftly transforming the field of mental healthcare by allowing for multi-institutional, privacy-preserving analytics that do not require the centralization of sensitive patient data. In this systematic review, conducted in accordance with PRISMA guidelines, we synthesise evidence from 20 studies focused on FL for mental state detection (including emotion recognition), as well as 25-30 studies encompassing broader behavioural health and human activity applications published up to December 2024. Our work maps FL research across a comprehensive taxonomy—encompassing FL paradigm, data modality, mental health condition, aggregation strategy, scale, and real-world evaluation metrics, revealing the breadth and limitations of FL adoption in stress, depression, bipolar disorder, OCD, schizophrenia, emotion recognition, and other domains. FL models, such as FedAvg, CAFed, FedHealth, and FedUFO, achieve high diagnostic accuracy (85–95 %) on leading public datasets (e.g., WESAD, DAIC-WOZ), with only minor trade-offs (2–5 % decrease in accuracy) compared to centralized learning FL methods. Boosting model generalizability and robustness by exploiting data diversity across sites, and expanding integrations with blockchain technology reinforce auditability and trustworthiness. We provide further support to mental health professionals and policymakers on workflow integration, data governance, auditability, and multi-center collaboration, as well as a roadmap for practical deployment that includes user-centered evaluation and regulatory compliance. This review provides the first taxonomy-driven, methodologically rigorous overview of FL in mental health, critically evaluating the field’s current state, highlighting actionable paths for real-world adoption, and identifying open research opportunities that will guide the next generation of secure and inclusive mental health technologies.