Scholarlink

Wireless Body Area Networks (WBANs) have emerged as potential game changers for continuous physiological monitoring, enabling informative healthcare interventions and telemedicine. Despite their huge advantages, WBAN data are prone to noise, abrupt sensor malfunctions, and various anomalies, leading to diminished reliability and effectiveness if left undetected. In this paper, we propose a novel two-stage anomaly detection model featuring a Convolutional Bidirectional LSTM (Conv-BiLSTM) as the backbone. Our framework first isolates point anomalies, such as sudden spikes in a single sample or sensor detachments, and then focuses on more intricate contextual anomalies with multiple signals. We validate our approach on a custom dataset collected from a controlled laboratory setting, and to assess generalizability in broader clinical contexts, we apply our model to the public MIMIC dataset. Our results consistently demonstrate high accuracy, recall, and low inference delay—key requirements for real-time healthcare monitoring. Through our model we have achieved 99.65% accuracy for Point anomaly (Stage 1) and 99.72% for contextual anomaly (Stage 2). The two-stage design of hybrid Conv-BiLSTM outperforms single-stage methods, highlighting the advantage of decoupling abrupt outliers from multi-sensor correlated anomalies in WBANs.

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

CNN-BiLSTM Based Deep Neural Architecture for Physiological Anomaly Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Shreea Bose,
  • Granth Bagadia,
  • Arundhati Bajaj,
  • Chittaranjan Hota

摘要

Wireless Body Area Networks (WBANs) have emerged as potential game changers for continuous physiological monitoring, enabling informative healthcare interventions and telemedicine. Despite their huge advantages, WBAN data are prone to noise, abrupt sensor malfunctions, and various anomalies, leading to diminished reliability and effectiveness if left undetected. In this paper, we propose a novel two-stage anomaly detection model featuring a Convolutional Bidirectional LSTM (Conv-BiLSTM) as the backbone. Our framework first isolates point anomalies, such as sudden spikes in a single sample or sensor detachments, and then focuses on more intricate contextual anomalies with multiple signals. We validate our approach on a custom dataset collected from a controlled laboratory setting, and to assess generalizability in broader clinical contexts, we apply our model to the public MIMIC dataset. Our results consistently demonstrate high accuracy, recall, and low inference delay—key requirements for real-time healthcare monitoring. Through our model we have achieved 99.65% accuracy for Point anomaly (Stage 1) and 99.72% for contextual anomaly (Stage 2). The two-stage design of hybrid Conv-BiLSTM outperforms single-stage methods, highlighting the advantage of decoupling abrupt outliers from multi-sensor correlated anomalies in WBANs.