Scholarlink

Machine learning (ML) techniques are crucial for improving diagnostic accuracy in psychiatry using neuroimaging-based biomarkers. Deep learning models like Kolmogorov Arnold Networks (KANs) are particularly promising in this context but struggle with high-dimensional datasets. We propose the Ensemble-KAN (E-KAN) method to overcome these limitations, integrating multiple base learners. Our novel approach aims to advance classification especially when multiple sources of data are available. The E-KAN was tested against traditional ML models in discriminating recent-onset psychosis (ROP) or depression (ROD) from healthy controls using multimodal environmental and neuroimaging data and it underwent a rigorous ablation study to test its effectiveness. Results demonstrate enhanced performance over traditional ML models, highlighting the efficacy of E-KAN models in psychiatric diagnostics. Specifically, our E-KAN achieved an accuracy of 72.5%, outperforming single-KAN models and traditional ML algorithms. This study underscores the potential of E-KAN models in advancing psychiatric research and personalized medicine through improved diagnostic capabilities. The code is available at https://github.com/brainpolislab/E-KAN .

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

Ensemble-KAN: Leveraging Kolmogorov Arnold Networks to Discriminate Individuals with Psychiatric Disorders from Controls

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Gianluca De Franceschi,
  • Inês W. Sampaio,
  • Stefan Borgwardt,
  • Joseph Kambeitz,
  • Lana Kambeitz-Ilankovic,
  • Eva Meisenzahl,
  • Raimo K. R. Salokangas,
  • Rachel Upthegrove,
  • Stephen J. Wood,
  • Nikolaos Koutsouleris,
  • Paolo Brambilla,
  • Eleonora Maggioni

摘要

Machine learning (ML) techniques are crucial for improving diagnostic accuracy in psychiatry using neuroimaging-based biomarkers. Deep learning models like Kolmogorov Arnold Networks (KANs) are particularly promising in this context but struggle with high-dimensional datasets. We propose the Ensemble-KAN (E-KAN) method to overcome these limitations, integrating multiple base learners. Our novel approach aims to advance classification especially when multiple sources of data are available. The E-KAN was tested against traditional ML models in discriminating recent-onset psychosis (ROP) or depression (ROD) from healthy controls using multimodal environmental and neuroimaging data and it underwent a rigorous ablation study to test its effectiveness. Results demonstrate enhanced performance over traditional ML models, highlighting the efficacy of E-KAN models in psychiatric diagnostics. Specifically, our E-KAN achieved an accuracy of 72.5%, outperforming single-KAN models and traditional ML algorithms. This study underscores the potential of E-KAN models in advancing psychiatric research and personalized medicine through improved diagnostic capabilities. The code is available at https://github.com/brainpolislab/E-KAN .