Anomaly Detection in WBANs using Semi-Supervised Double Deep Q-Networks

The Wireless Body Area Network (WBAN) seamlessly interconnects wearable devices on the human body, enabling the monitoring of diverse physiological parameters for better health outcomes. Ensuring the accuracy and security of data transmission is essential to avert the occurrence of false alarms and maintain the overall reliability of WBANs. The substantial volume of data exchanged within WBANs heightens the vulnerability to data attacks or the transmission of noisy or erroneous data, necessitating real-time handling capabilities. Moreover, the availability of labeled datasets is limited. Consequently, this research proposes a semi-supervised Double Deep Q-Network (DDQN) model capable of real-time detection of false alarms within the data. This model incorporates a DDQN featuring a Deep Neural Network architecture that includes an LSTM layer and several Dense layers to capture the intricate spatiotemporal relationships within the dataset and classify the data received into normal or faulty categorical data. The semi-supervised DDQN model leverages a K-Nearest Neighbor (KNN) classifier to assign labels to the unlabeled data extracted from the MIMIC-1 dataset. Subsequently, this labeled data is forwarded to the DDQN model for prediction. To validate the efficacy of the semi-supervised DDQN model, we have performed extensive testing, resulting in an accuracy of 98.9%, a recall of 96.48%, a precision of 92.25%, an F1 score of 94.32%, and a false-positive rate of 0.85%.