Scholarlink

<p>Wearable sensors and biosensors are significantly used in the field of medicine. There are numerous classification algorithms available in the literature to categorize sensor signals on the basis of parameters including accuracy, signal reliability, memory footprints and processing efficiency; this includes convolutional neural networks (CNN) and recurrent neural networks (RNN). To fill these gaps, the current research uses multimodal fusion with a shared-latent mapping system to enhance the accuracy of classification, signal stability and processing efficiency of wearable biosensor systems in the missing or degraded data conditions. Simulations were performed by expending nonlinear device metrics derived from wearable signals and IoT-cloud combined data. The key parameters measured included the covert dimension, paired data ratio, noise reduction, inference time, memory footprint, latency, energy, and throughput. The proposed system provides an increased latent dimension from 64 to 128, an unpaired training strategy eliminating 35% of the paired data requirement, and a combined denoising and adversarial objective, intending to improve the SSIM from 0.72 to 0.89, the PSNR from 25.6 to 31.8&#xa0;dB, and noise reduction from 5.2 to 24.6&#xa0;dB while enhancing pixel sharpness from 12 to 32.1%. In addition, the proposed multimodal model achieved an accuracy of 98.6%, recall of 97.9%, F1-score of 98.2%, and 10.4&#xa0;ms latency. The results obtained from this proposed work validate the important improvements over the existing method with respect to performance parameters such as accuracy, stability, and computational efficiency.</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

Deep learning framework for multi-modal biosensor fusion in wearable health monitoring systems

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月26日

查看全文
  • B. Sabitha,
  • K. Akila,
  • R. Saravanan,
  • M. Mathankumar

摘要

Wearable sensors and biosensors are significantly used in the field of medicine. There are numerous classification algorithms available in the literature to categorize sensor signals on the basis of parameters including accuracy, signal reliability, memory footprints and processing efficiency; this includes convolutional neural networks (CNN) and recurrent neural networks (RNN). To fill these gaps, the current research uses multimodal fusion with a shared-latent mapping system to enhance the accuracy of classification, signal stability and processing efficiency of wearable biosensor systems in the missing or degraded data conditions. Simulations were performed by expending nonlinear device metrics derived from wearable signals and IoT-cloud combined data. The key parameters measured included the covert dimension, paired data ratio, noise reduction, inference time, memory footprint, latency, energy, and throughput. The proposed system provides an increased latent dimension from 64 to 128, an unpaired training strategy eliminating 35% of the paired data requirement, and a combined denoising and adversarial objective, intending to improve the SSIM from 0.72 to 0.89, the PSNR from 25.6 to 31.8 dB, and noise reduction from 5.2 to 24.6 dB while enhancing pixel sharpness from 12 to 32.1%. In addition, the proposed multimodal model achieved an accuracy of 98.6%, recall of 97.9%, F1-score of 98.2%, and 10.4 ms latency. The results obtained from this proposed work validate the important improvements over the existing method with respect to performance parameters such as accuracy, stability, and computational efficiency.