DeepCAD: A Stand-alone Deep Neural Network-based Framework for Classification and Anomaly Detection in Smart Healthcare Systems

Abstract

Contemporary smart healthcare systems (SHSs) frequently use wireless body sensor devices (WBSDs) for vital sign monitoring and the internet of medical things (IoMT) network for rapid communication with a cloud-based controller. The SHS controllers generate required control decisions based on the patient status to enable real-time patient medication/treatment. Hence, the correct medical delivery primarily depends on accurately identifying the patient's status. Accordingly, SHSs mostly leverage deep neural network (DNN)-based machine learning (ML) models for patient status classification due to their prediction accuracy and complex relation capturing capability. Nevertheless, the open IoMT network is prone to several cyberattacks, including adversarial ML-based attacks, which can exploit DNN models and create a life-threatening event in a safety-critical SHS. Existing solutions usually propose outlier detection or transfer learning-based ML models on top of the patient status classification model to deal with SHS security issues. However, incorporating a separate anomaly detection model increases the model complexity and raises feasibility issues for real-time deployment. This work presents a novel framework, DeepCAD, that considers training a stand-alone DNN model integrated with anomaly detection rules for classification and anomaly detection in SHS. The proposed framework is verified on the Pima Indians Diabetes and Parkinson datasets.