Scholarlink

<p>Continuous monitoring of physiological signals using wearable devices enables early detection of health anomalies; however, existing learning-based approaches often struggle with subject-specific variability, missing data, and limited temporal modeling, resulting in reduced robustness and elevated false-positive rates. To address these challenges, this study proposes PerGen-ATNet, a novel personalized generative–attention pipeline for anomaly detection in multimodal wearable health signals. Unlike existing methods that rely on generic fusion and population-level modeling, the proposed framework introduces an individualized latent learning mechanism that captures subject-specific physiological patterns. The model employs a variational transformer-based architecture that jointly integrates modality-specific encoding, cross-sensor attention, and probabilistic latent representation within a unified pipeline. Heterogeneous signals, including ECG, PPG, SpO<sub>2</sub>, and accelerometer data, are first mapped into a shared personalized latent space, where transformer attention explicitly models temporal dependencies and inter-modality relationships. A key contribution lies in the personalized latent regularization, enabling the model to learn adaptive health baselines rather than fixed distributions. The generative decoder reconstructs expected normal patterns, and anomalies are detected using a dual-criterion strategy combining reconstruction error and latent distribution deviation. Evaluation is conducted on Three publicly available and clinically validated datasets, including MIMIC-III, BIDMC, and WESAD, using stratified cross-validation to ensure statistically reliable performance estimates. The proposed PerGen-ATNet framework achieves superior anomaly detection performance, attaining an average accuracy of 96.8%, precision of 95.9%, recall of 96.3%, F1-score of 96.1%, and AUC of 0.982 across the MIMIC-III, BIDMC, and WESAD datasets, consistently outperforming CNN-, RNN-, and transformer-based baselines by a margin of 3–6% in key evaluation metrics while maintaining low false-positive rates and robust cross-subject generalization.</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

Personalized Generative Attention Modeling for Robust Anomaly Detection and Real-Time Monitoring in Multimodal Wearable Health Signals

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Ramya Srikanteswara,
  • Suchi Mishra,
  • Ch. Subbalakshmi,
  • Sachin Tiwari,
  • Naglaa F. Soliman,
  • Mohamed M. Hassan

摘要

Continuous monitoring of physiological signals using wearable devices enables early detection of health anomalies; however, existing learning-based approaches often struggle with subject-specific variability, missing data, and limited temporal modeling, resulting in reduced robustness and elevated false-positive rates. To address these challenges, this study proposes PerGen-ATNet, a novel personalized generative–attention pipeline for anomaly detection in multimodal wearable health signals. Unlike existing methods that rely on generic fusion and population-level modeling, the proposed framework introduces an individualized latent learning mechanism that captures subject-specific physiological patterns. The model employs a variational transformer-based architecture that jointly integrates modality-specific encoding, cross-sensor attention, and probabilistic latent representation within a unified pipeline. Heterogeneous signals, including ECG, PPG, SpO2, and accelerometer data, are first mapped into a shared personalized latent space, where transformer attention explicitly models temporal dependencies and inter-modality relationships. A key contribution lies in the personalized latent regularization, enabling the model to learn adaptive health baselines rather than fixed distributions. The generative decoder reconstructs expected normal patterns, and anomalies are detected using a dual-criterion strategy combining reconstruction error and latent distribution deviation. Evaluation is conducted on Three publicly available and clinically validated datasets, including MIMIC-III, BIDMC, and WESAD, using stratified cross-validation to ensure statistically reliable performance estimates. The proposed PerGen-ATNet framework achieves superior anomaly detection performance, attaining an average accuracy of 96.8%, precision of 95.9%, recall of 96.3%, F1-score of 96.1%, and AUC of 0.982 across the MIMIC-III, BIDMC, and WESAD datasets, consistently outperforming CNN-, RNN-, and transformer-based baselines by a margin of 3–6% in key evaluation metrics while maintaining low false-positive rates and robust cross-subject generalization.