Energy-efficient adaptable implantable body sensor networks

Abstract

During the last decades, wireless technologies have revolutionized the healthcare industry by enabling remote monitoring of physiological signals of the human body through wearable medical sensors connected to the Internet. Currently new medical devices, such as the implantable deep-brain stimulator, are emerging that not only sense the physiological signals but also can deliver targeted disease-specific medical therapy. While traditionally many implantable medical devices have operated in isolation, recently many attempts have been made to connect them and to form a network, the so-called Implantable Body Sensor Networks (IBSN). In its basic form, IBSN should enable medical devices (both implantable and wearable) to communicate with each other, in and around the patient's body, so that a common consensus on the health-condition of the patient can be reached. This collective knowledge will enable predictive diagnostics for the patient and will be used to deliver a personalized medical therapy based on patient’s physiological status. One of the important challenges of wireless communication in IBSN is that human body, due to its conductive property, heavily attenuates the low-power radio frequency (RF) signals. Due to IBSN node’s utmost proximity to the human tissues, transmission power cannot be increased to compensate for the attenuation. Therefore, even a minor human movement frequently disrupts the RF link quality between the nodes. In addition, sensor nodes of IBSN are heterogeneous, i.e. they can sense and/or deliver therapy, generate and/or receive a wide variety of data at the nodes, which requires an adaptive Quality of Service (QoS) to support this heterogeneity. Nodes of IBSN are heavily constrained in power, computation, and memory resources given their size and placement, yet they are expected to operate for a long-time without needing any maintenance. Moreover these nodes must have ultra-low power consumption and very high reliability to ensure their long-term and life-critical functionalities. In this thesis, we primarily focus on developing energy-efficient and reliable wireless communication schemes for resource-constrained nodes of IBSN that are adaptable to the requirements of variety of medical applications. To this end, we first measure and model the physical wireless channel in and around the body using an animal tissue that closely resembles the human tissue. With this model, we analyze the impact of attenuation caused by the body tissue to the wireless channel by simulating the state-of-the-art network protocols and measuring various network parameters such as throughput and latency at different locations in and around the tissue and different distances between the implantable nodes. Based on these results and the requirements of advanced medical devices, we develop a cross-layer optimized medium access control protocol called HACMAC (Human ACtivity based Medium Access Control), which is tolerant to frequent link disruptions. The HACMAC protocol uses simple statistical measures to create a RF disruption model for each human activity. Based on this model, it will compute the optimal duty-cycle for the wireless communication and adjust the Medium Access Control (MAC) layer parameters. However, the RF disruption model, which is developed based on simple statistical measures, cannot cope up with the minor yet frequent disruption caused by variations in human activities. To cope up with this dynamically changing human activity based RF patterns, we use reinforcement-learning principles to further improve the reliability of wireless communication and present a MAC optimization framework called DiNAMAC that is independent of RF disruption-models. Also, we use wake-up radio in addition to the main radio to lower down the overall power consumption of heterogeneous medical devices without affecting the reliability of the wireless communication. Finally, we develop a context-aware seizure prediction framework for an epileptic use-case, which reduces the amount of raw data transmission by locally processing the brain signals on the IBSN sensor node. In addition, we investigate low-power low-complex signal processing algorithms and machine-learning models capable of being executed on the resource constrained sensor nodes. Based on these models, we develop an active learning-based seizure prediction framework and compare its performance with conventional methods of seizure prediction. To further analyze the feasibility of our optimization framework and seizure prediction framework to be executed on a resource constrained sensor node, we implement the core of our framework on an ARM based Cortex M4 microprocessor to measure the resource consumption values. From our results, it is shown that it is not only feasible to execute our framework on an IBSN sensor node, but also to reduce the overall power consumption of the node without affecting the QoS parameters of the wireless communication.