A Convolutional Transformer Network for Anomaly Detection in Wireless Body Area Networks

The Wireless Body Area Network (WBAN) integrates wearable devices and IoT sensors in the human body, enabling real-time monitoring of physiological parameters for improved healthcare outcomes. Ensuring accurate and reliable data transmission is crucial to maintain system performance. To address this, we propose a novel anomaly detection framework that uses a two-stage ConvTransformer architecture, specifically designed to handle both point anomalies (isolated abnormal readings) and contextual anomalies (irregularities across multiple signals simultaneously) in physiological data. The model was optimized to handle real-time data processing, ensuring quick response times in critical healthcare scenarios. In the first stage, we trained a ConvTransformer model to distinguish between humane data and point anomalies. These random out-of-range values may indicate abrupt irregularities in individual sensor readings. After identifying and filtering out point anomalies, the second stage applies another ConvTransformer model to the remaining humane data to detect contextual anomalies. These are more complex and involve simultaneous irregularities in multiple physiological signals (for example, heart rate, body temperature, SpO $_{2}$ , and ECG), which may suggest more significant health concerns. We developed a custom dataset in our laboratory using IoT sensors to simulate realistic health scenarios. This two-stage detection approach ensures more precise and robust anomaly detection by first eliminating apparent point anomalies and then focusing on the more intricate patterns of contextual anomalies. The first model achieved 99.66% accuracy in detecting point anomalies, while the second model reached nearly 99.76% accuracy in identifying contextual anomalies, showcasing the efficiency of the ConvTransformer architecture in WBAN applications for detecting anomalies.