Scholarlink

<p>Eye rubbing constitutes a primary modifiable risk factor for keratoconus progression and post-refractive surgery ectasia. Current automated detection systems using single-paradigm architectures exhibit suboptimal performance in capturing complex spatiotemporal motion signatures, while facing significant computational constraints on resource-limited wearable devices. In this paper, we propose a hybrid CNN-Transformer architecture for eye-rubbing detection using inertial measurement unit sensor data from consumer smartwatches, addressing the critical challenge of real-time processing on edge devices with limited computational resources. We developed a smartwatch-based application that evaluates multiple deep learning architectures, including gated recurrent units, long-short-term memory networks, convolutional neural networks (CNNs), Transformers, and hybrid models, to analyze time-series sensor data, while optimizing for computational efficiency and real-time performance. Our approach uses signal segmentation to divide data into fixed-length segments, processed through a one-dimensional (1D) CNN with Transformer-based self-attention to extract temporal and spatial features. The proposed CNN-Transformer architecture achieved 99.92% accuracy, 99.93% sensitivity, and 99.97% specificity, while maintaining real-time inference latency below 50ms on resource-constrained smartwatches, outperforming existing methods that report 90.3–97% accuracy with significantly higher computational overhead. Statistical analysis confirmed significant improvements (<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11227_2025_7744_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="76" /> </InlineMediaObject> <EquationSource Format="TEX">\(p\,&lt;\,0.001\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>p</mi> <mspace width="0.166667em" /> <mo>&lt;</mo> <mspace width="0.166667em" /> <mn>0.001</mn> </mrow> </math></EquationSource> </InlineEquation>) across evaluation metrics.</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

Wearable sensor-based eye-rubbing monitoring: a hybrid CNN-attentionrub architecture for keratoconus prevention

  • 期刊论文

  • 发表日期：

    2025年8月23日

查看全文
  • Ayoub Louja,
  • Ines Drira,
  • Abdellah Jamali,
  • Najib Naja,
  • Layth Sliman

摘要

Eye rubbing constitutes a primary modifiable risk factor for keratoconus progression and post-refractive surgery ectasia. Current automated detection systems using single-paradigm architectures exhibit suboptimal performance in capturing complex spatiotemporal motion signatures, while facing significant computational constraints on resource-limited wearable devices. In this paper, we propose a hybrid CNN-Transformer architecture for eye-rubbing detection using inertial measurement unit sensor data from consumer smartwatches, addressing the critical challenge of real-time processing on edge devices with limited computational resources. We developed a smartwatch-based application that evaluates multiple deep learning architectures, including gated recurrent units, long-short-term memory networks, convolutional neural networks (CNNs), Transformers, and hybrid models, to analyze time-series sensor data, while optimizing for computational efficiency and real-time performance. Our approach uses signal segmentation to divide data into fixed-length segments, processed through a one-dimensional (1D) CNN with Transformer-based self-attention to extract temporal and spatial features. The proposed CNN-Transformer architecture achieved 99.92% accuracy, 99.93% sensitivity, and 99.97% specificity, while maintaining real-time inference latency below 50ms on resource-constrained smartwatches, outperforming existing methods that report 90.3–97% accuracy with significantly higher computational overhead. Statistical analysis confirmed significant improvements ( \(p\,<\,0.001\) p < 0.001 ) across evaluation metrics.