Scholarlink

Background <p>Digital interventions for mental health promotion are highly scalable, yet often remain generic. Data-driven personalization based on intensive longitudinal data from everyday life, so-called ecological momentary assessments (EMA), could enable context-sensitive tailoring of interventions but requires demonstrated feasibility under real-world conditions in terms of technical, user-related and modeling-related aspects.</p> Objective <p>The feasibility of an adaptive, artificial intelligence (AI)-based personalization system for assignment of outpatient mobile health (mHealth) interventions (ecological momentary interventions, EMI) in the living laboratory AI4U was evaluated.</p> Material and methods <p>In three microrandomized controlled studies EMAs were collated over 40 days and assigned to EMIs. For every individual (<i>n</i> = 145) a&#xa0;person-specific time-series model was trained daily, which generated EMI suggestions. Primary feasibility criteria were technical stability, EMA compliance and suitability of the time-series data for individualized modeling.</p> Results <p>The median EMA compliance was approximately 50% across the 40-day period. About half of the EMI allocations were carried out algorithmically or randomized by the server; errors were compensated by a&#xa0;failover mechanism integrated in the app. The time-series exhibited sufficient variance and pronounced interindividual heterogeneity. As the amount of data increased, the person-specific models converged, indicating increased reliability of the model parameters.</p> Conclusion <p>Data-driven personalization of ambulatory interventions is principally, technically and methodologically feasible under real-life conditions. Further optimization is needed especially regarding infrastructure and long-term compliance. Adaptive time-series models offer a&#xa0;controllable and interpretable alternative to generative AI systems.</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

KI-basierte Personalisierung digitaler Mikrointerventionen

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月16日

查看全文
  • Janik Fechtelpeter,
  • Christian Rauschenberg,
  • Christian Götzl,
  • Selina Hiller,
  • Eva Wierzba,
  • Silvia Krumm,
  • Ulrich Reininghaus,
  • Daniel Durstewitz,
  • Georgia Koppe

摘要

Background

Digital interventions for mental health promotion are highly scalable, yet often remain generic. Data-driven personalization based on intensive longitudinal data from everyday life, so-called ecological momentary assessments (EMA), could enable context-sensitive tailoring of interventions but requires demonstrated feasibility under real-world conditions in terms of technical, user-related and modeling-related aspects.

Objective

The feasibility of an adaptive, artificial intelligence (AI)-based personalization system for assignment of outpatient mobile health (mHealth) interventions (ecological momentary interventions, EMI) in the living laboratory AI4U was evaluated.

Material and methods

In three microrandomized controlled studies EMAs were collated over 40 days and assigned to EMIs. For every individual (n = 145) a person-specific time-series model was trained daily, which generated EMI suggestions. Primary feasibility criteria were technical stability, EMA compliance and suitability of the time-series data for individualized modeling.

Results

The median EMA compliance was approximately 50% across the 40-day period. About half of the EMI allocations were carried out algorithmically or randomized by the server; errors were compensated by a failover mechanism integrated in the app. The time-series exhibited sufficient variance and pronounced interindividual heterogeneity. As the amount of data increased, the person-specific models converged, indicating increased reliability of the model parameters.

Conclusion

Data-driven personalization of ambulatory interventions is principally, technically and methodologically feasible under real-life conditions. Further optimization is needed especially regarding infrastructure and long-term compliance. Adaptive time-series models offer a controllable and interpretable alternative to generative AI systems.