Developing and enabling edge analytics for Internet of Medical Things
This project seeks to investigate and develop robust and resource-based edged algorithms for applications in the Internet of Things and the development of intelligent applications of real-time data-enhanced solutions in the e-Health domains.
Machine learning (ML) and deep learning (DL) are promising tools for developing intelligent applications for various domains, especially in the e-health domain. Unfortunately, these algorithms are typically very resource-intensive and require a lot of computational resources to deploy and test. The Internet of Medical Things (IoMT) is also a resource-constrained ecosystem, despite its evolution and applications in recent times. Meanwhile, due to the ground-breaking deployment of 5G technology, the challenge of resource constraint seems to be solvable as new improved algorithms seek to improve connectivity, cloud-based storage, and extend the computability of machines while expanding the applicability of mobile IoMT (West, 2016; Patel et al., 2017). Evidently, there are emerging applications of edge IoMT devices in care delivery, as connected medicine seems to provide advanced dynamics for obtaining quality care through imaging improvement, diagnostics, and treatment of all types of complex health challenges.
Therefore, this project will be focused on building innovative solutions to design compact and lightweight machine learning and deep learning algorithms to fit, train, and deploy these algorithms on resource-constrained edge devices for tracking, early detection, and diagnosis of diseases without compromising their performance and accuracy. Considering the nature of tracking that is required for the device to perform and the eventual detection or prediction of the possible presence of diseases in the body, distributed and federated learning techniques will be used over multi-modal data sets (e.g. sensor data streams, historical datasets) to provide robust and accurate Artificial Intelligence (AI) models for application in tracking, detecting, and possible enhancement of the different chemotherapy stages for affected patients.
To deploy edged IoMT devices, such as clinical wearables and remote sensors, as well as many other devices that monitor and electronically transmit medical data, such as vital signs, physical activity, personal safety, and medication adherence, the project specifically aims to develop robust algorithms based on big data analytics, learning, and intelligence. Smart objects are now the ideal building blocks for the creation of cyber-physical smart ubiquitous frameworks thanks to the IoMT. With bright technological, economic, and societal prospects, the IoMT revolution is redefining
contemporary healthcare (Islam et al., 2015). These tools will offer diagnosis and treatment options in telemedicine that have never been seen before, all while providing high-quality care at a reasonable cost for patients.
In summary, this project seeks to explore the use of knowledge graph generation and explainable AI techniques over IoMT data for the tracking, detection, and diagnosis of diseases in the human body.
Validation and Quantification of optical properties of water using machine learning
Optical properties of water provide information about the condition of water bodies which can also give information on water pollution. Various water quality parameters such as the total suspended matter (TSM), chlorophyll and colour dissolved organic matter (CDOM) concentrations were traditionally based on estimations from in situ data. This method is expensive, time-consuming and provided local coverage. However, technology has been progressively improving and advancing. More remote sensing data and platforms are now freely and easily accessible. Water quality can now be extracted from the satellite imageries with relatively good spatial resolution, data is more frequently collected at a global coverage and at a good accuracy. A strong positive correlation coexists between the in situ data and satellite imageries on water properties. Though, machine learning can assist in further improving the accuracy and automate the process by using trained models. The size of dataset tends to influence the accuracy of remote sensing results. Training a model in machine learning requires a large dataset which eventually improves the accuracy of remote sensing. The process mainly involves preparing the data for training in a model. A suitable machine learning model is selected and model is fit into the data. It involves learning the pattern of the data and using the pattern to make predictions. Next, the model is evaluated to assess the accuracy and precision of the model. If the accuracy of the model is low, tuning and experimentation is carried out. Which improves the accuracy of the model by manipulating the hyperparameters and comparing accuracy of the model from the various hyperparameters used. The model is saved once the required accuracy is acquired. This can then be reloaded and reused for making predictions and classification. Which can assist in making prediction on water quality parameters, classification of satellite imageries into respective feature classes or detect specific features in an imagery. Thus, machine learning assists in predictions of both in situ and satellite data, classification and feature detection which ultimately saves time, money and improves the accuracy. However, various challenges are expected in this research project. Such as finding satellite imageries with good spatial resolution and with less than 10 percent of cloud cover. Moreover, training the model and improving its accuracy can be time-consuming. Despite the challenges, the research project will provide an insight into the potential of remote sensing and machine learning in extracting the water optical properties
Interpretability and Visualization of Deep Learning Models in Magnetic Resonance Imaging and Radiology
Deep learning approaches are achieving state-of-the-art results in Magnetic Resonance Imaging (MRI) and Radiology (Computed Tomography, X-ray) on such downstream tasks as image classification and image segmentation. However, they still suffer from a lack of human interpretability critical for increasing understanding of the methods’ operation, enabling clinical translation and gaining clinician trust. Therefore, it is crucial to provide clear and interpretable rationale for model decisions.
In order to understand model predictions, there are algorithms that can be applied to a fully trained deep learning model and produce heatmaps highlighting portions of the image that contributed to the model predictions. These algorithms are (but are not limited to) Guided-Backpropagation, Class Activation Mapping (CAM), Grad-CAM, Grad-CAM++ and Score-CAM.
The recent COVID‐19 pandemic made clear that rapid clinical translation of Artificial Intelligence (AI) systems may be critically important for assisting treatment‐related decisions and improving patient outcomes. Recently, there has been a particular interest in applying the above algorithms to interpret COVID-19 Chest X-rays and assist clinicians in making the correct diagnosis.
We will be working with the open source MRI, CT and X-ray public datasets and will initially investigate:
How different Convolutional Neural Networks (CNNs) architectures contribute to the generation of reliable heatmaps. Here, we will focus on both custom CNNs and established pre-trained transfer learning based architectures such as VGG, ResNet, Inception, etc.
Weight initialization in CNNs has been also shown (and is in line with our experience) to be important to the performance of the interpretability algorithms and we will investigate different weight initialization protocols. We will attempt to answer the research question why the initialization with ImageNet weights that are completely unrelated to the medical imaging domain leads to the astonishing performance of some interpretability algorithms on medical imaging datasets.
Most works in the literature are focused on the interpretability methods for the classification models but there are some limited works that also examine the interpretability of segmentation networks. We will examine both downstream tasks and for this purpose we will be working with the 2D T1-weighted CE-MRI dataset compiled by Cheng et al. as this dataset contains both classification and segmentation labels.
Therefore, this study will also consider weakly supervised image segmentation of the Cheng et al. dataset using class-specific heatmaps, possibly Conditional Random Fields (CRF), and thresholding techniques.
Many interpretability algorithms are applied to the last convolutional layer of CNNs and therefore the resulting heatmaps are coarse. We will investigate how to leverage the intermediate layers for the generation of finer heatmaps.
To summarize, this project is expected to identify best performing approaches for providing explainable and interpretable AI output and discuss their advantages and disadvantages for the MRI and Radiology datasets considered in this study. Further, we expect to provide recommendations for appropriate incorporation of these techniques to improve deep learning models and evaluate their performance.
Investigating The Use of Machine Learning in Health Science
The benefits in regular physical activity (PA) participation are well documented with irrefutable evidence of the effectiveness of regular PA in the prevention of several chronic diseases (e.g., cardiovascular disease, diabetes, obesity, depression, etc.) and premature death (Warburton et al., 2006). In Ireland, the Children’s Sport Participation and Physical Activity (CSPPA) study found just 17% of Irish children engaged in the recommended one hour per day of moderate to vigorous physical activity, which is a drop from the 19% recorded in 2010. (Woods et al., 2018). Therefore, it seems important to explore factors that influence PA participation and a means to measure impact.
Machine learning (ML) has become a common way to measure physical activity (Narayanan et al., 2020). Accelerometer data processing techniques based on pattern recognition have been shown to provide accurate predictions of physical activity type and more accurate assessments of physical activity intensity (Trost et al., 2012, 2018; Ellis et al., 2016). Nonetheless, the uptake of machine learning methods by physical activity researchers has been slow, in part due to the difficulties of implementation, and the consistent finding that models trained on accelerometer data from laboratory-based activity trials do not generalize well to free-living environments (Sasaki et al., 2016; Lyden et al., 2014; Bastian et al., 2015). Recently, it has been argued that ML has failed physical activity research in four important ways: a lack of benchmark data, priority in methods development, limited software integration and absence of training (Fuller, Ferber and Stanley, 2022). To improve the use of ML methods in physical activity research, it has been proposed that as a discipline, practitioners must use and publish benchmark datasets to allow for increased opensource methods development. Many datasets currently exist on health-related fitness, wellbeing, confidence and motivation towards PA amongst the Irish youth demographic, a recent example being the Moving Well-Being Well project, a national study which assessed a range of variables linked to PA participation in over 2,000 Irish children. One finding from the study highlighted the low levels of fundamental movement skills (FMS) mastery in Irish primary school children (Behan et al., 2019). FMS proficiency has shown positive associations with increased physical activity in both children and adolescents (Barnett et al., 2009; Lubans et al., 2010).
Current PA-related datasets are limited in terms of the analyses undertaken, due to a multitude of factors, a lack of training being one of them (Fuller, Ferber and Stanley, 2022). Given the size and complexities of these datasets, as well as the societal need to create a healthier Ireland, it would appear necessary that relevant ML techniques are explored in an effort to uncover further unknown patterns and correlations/disparities that exist between specific age groups and genders. This research aims to explore the use of ML in health science, specifically focusing upon children and youth with a further aim of utilizing ML in an appropriate manner towards PA to have a societal impact on Irish health.
Proactive Machine Learning-based Methods for Intelligent and Secure Vehicular Networks
Vehicular networks have a high potential in the creation of smarter cities, but also smarter roads. This potential relies on the on the wheels connectivity provided by vehicular networks that can also meet the always connected need of drivers and passengers as they are spending much of their daily time in their vehicles. Moreover, the vehicular network is considered to have a crucial role in the context of self-driving vehicles. Vehicular networks are based on “smart” vehicles that are able to communicate to each other and to the infrastructure via vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communications, known under the generic term of V2X communications, but also via other wireless communications technologies.
Security is an important concern in vehicular networks as malicious attacks can lead not only to material losses, or reputation loss, etc., but can actually lead to the loss of human lives. Vehicular networks have specific characteristics that pose challenges to security, such as high mobility, rapidly changing topology, large scale, heterogeneous environment, etc. The implementation of security mechanisms in vehicular networks usually introduce an overhead in communication. In particular, the proactive mechanisms lead to high signalling and communication cost. Hence, the state of the art is represented by reactive detection in order to balance the quality of service and security. Machine learning (ML) algorithms have been used successfully in network security in general, and vehicular networks in particular. In the latter case, they demonstrated their potential of dealing with the challenges imposed by the aforementioned unique characteristics of vehicular networks.
This project’s main research objectives are listed below, but the ultimate goal of the project is to propose proactive security methods for vehicular networks.
RO1 – comprehensive literature review in the area of ML-based security mechanisms for vehicular networks.
RO2 – propose proactive security methods for vehicular networks that are based on ML. Blockchain will be also investigated as it is showing promises in this area.
Exploiting Rich Long-term Memory Context in Spatiotemporal Vision Tasks without Annotations
Videos are one of the richest and vastly produced data types. Due to its higher dimensionality and complexity, it is challenging to model spatiotemporal vision tasks with classical approaches. Recent research has focused on solving these tasks with deep learning techniques. Supervised learning has been the most successful version of these techniques, but it requires annotations to train the model. As video data is massively growing with time, it is impossible to create labels for different tasks to capture a wide variety of patterns. The research aims to solve this annotation scarcity in spatiotemporal vision tasks with self-supervised and unsupervised learning paradigms. In order to exploit memory context, recent research has used various internal memory modules. However, these internal memory modules do not exploit rich long-term memory elements. The research aims to solve this issue by introducing memory networks. Memory networks contain an external memory module synchronized with deep learning architecture. Unlike internal memory modules, they capture rich long-term past knowledge. On the other side, excellent qualitative results can also be obtained from these memory modules, proving their transparency and explainability. For achieving both aims, this proposal suggests testing various hypotheses obtained from memory networks and related literature on three different spatiotemporal vision tasks: (i) self-supervised video object segmentation (Self-VOS), (ii) video prediction, and (iii) unsupervised video anomaly detection.
Machine learning based characterisation of nanoparticles synthesized using microfluidics
The application of machine learning (ML) to microfluidics is currently gaining significant momentum within the research community particularly for applications in material science, nanomedicine and biologics. Microfluidics is the miniaturisation and automation of laboratory processes resulting in improved quality of data generated with reduced reagent and support costs. While most widely applied to diagnostic testing (e.g. pregnancy tests and COVID-19 testing). Microfluidics is also applied to high-throughput applications such as bioprocess monitoring, cell analysis, genetic screening, and drug discovery. These high-throughput applications involve large data sets of complex biological processes – thereby, making them ideally suited to control and optimisation through ML.
An emerging application of high-throughput microfluidics, particularly for drug discovery and in the development of drug delivery systems, is the use of microfluidics for nanoparticle synthesis. Furthermore, these same microfluidic platforms can also be used to investigate the interaction of nanoparticles with cells. These platforms can help optimise drug design and uptake, and therefore potentially improve patient outcomes for many chronic diseases (e.g., various types of cancer, HIV and diabetes).
The premise of the proposed research is based on ML application to microfluidic nanoparticle synthesis to optimise their desired properties (i.e., particle size and charge, polydispersity index, drug loading capacity and cell encapsulation). Then application of similar algorithms to optimise the nanoparticle interaction with cells for efficacious drug delivery and therapeutic response. In this project, the specific aim will be to implement different passive microfluidic techniques (e.g., hydrodynamic flow focusing, microvortices, chaotic advection and droplets generation) for efficient fluid mixing leading towards optimal nanoparticle synthesis. ML will be implemented through the development of at-line monitoring technology to measure the physical properties of the nanoparticles within the microfluidic channel. The coupling of this multimodal monitoring to ML is novel and can lead to innovative advances in the field of nanomedicine, biologics and organ-on-a-chip.
Privacy Preserving Federated Learning based Cognitive Digital Twins for Smart Cities
Recent advancements in AIoT (Artificial Intelligence + Internet of Things) techniques have made possible continuous advancements of various smart city applications in various sectors. Cognitive Digital Twin (CDT) – a Knowledge Graph and AI based virtual replica of the physical world, has been adopted by various industries especially the manufacturing sector but has been significantly slow adoption for smart city. The major reason being – lack of trust and privacy concerns towards sharing sensitive data. Privacy Preserving Federated Learning (PPFL), could be integrated along with CDT to ensure privacy preservation and trustworthiness. This research proposes a framework for integrating PPFL and CDT technologies to address various real life smart city scenarios as well as enable feasibility for smart city governance. A CDT is an extended or augmented version of DT with cognitive capabilities. It contains at least the three basic elements of DT – the physical entity (systems, subsystems, components etc.), digital (or virtual) representation or shadows and the connections between the virtual and physical spaces. The main difference is that CDT usually contains multiple DT models with unified semantics topology definitions. In addition, a CDT should – have cognition capabilities, have a digital version of the entire lifecycle of a system, be autonomous and be able to continuously evolve with the physical system across the entire lifecycle. The core idea of Federated Learning (FL) is to train machine learning models on datasets that are distributed across different devices or parties, which can preserve the local data privacy. Privacy Preserving Federated Learning (PPFL) focuses on the privacy preserving mechanism of FL by employing privacy preserving techniques such as Homomorphic Encryption (HE), Secure Multi-party Computation (SMC) and Differential Privacy (DP).
The motivation behind coupling PPFL and CDT are quite a few. Even though CDT has its own benefits, it lacks maintaining data privacy as data has to be transferred to the CDT from various sensors and physical entities. In PPFL, the privacy is preserved at the end user as only weights and parameters updates are transmitted from the edge devices and not the data. Hence raw sensitive information stays with the device and is not communicated over the network. The communications between smart city DTs as well as between DTs and edge devices is privacy-preserved. Last but not least, data quality and integrity is improved as pre-processing is done at the edge environment.
Research Objectives and Directions:
Create Ontology and Knowledge Engineering models for smart cities (serving as the data layer).
Research on creating a GNN based framework (or other enabling technology) to form the cognition capability of the CDT framework basis graph data models catering to smart cities.
Research on a framework creation of Privacy Preserving FL in general and specialise it for adoption of CDT.