Open Research Exeter - University of Exeter
Not a member yet
41213 research outputs found
Sort by
Ten-year analysis of Abbott FreeStyle Libre: a meta-analysis of continuous glucose monitoring accuracy
No full textBackground and Aims: Since its market launch in 2014, the Abbott FreeStyle Libre system has played a significant role in continuous glucose monitoring (CGM). This meta-analysis reviews a decade’s worth of research, examining the system’s accuracy across multiple clinical scenarios. The analysis com-pares Libre’s performance with that of other CGMs—particularly Dexcom devices—and evaluates its alignment with standard clinical reference methods. Methods: Published literature from 2014 to 2025 was systematically collected and categorized into six focus areas: (1) Head-to-head comparisons between Libre and Dexcom CGMs, (2) Libre’s standalone accuracy, (3) Clinical outcomes related to diabetes management, (4) User experience and system usability, (5) Evaluation of glucose trend arrows, and (6) Accuracy during exercise. Data from 15 studies were combined using a random effects model to account for differences in study design and populations. Results: The pooled analysis demonstrated no statistically significant difference in overall measurement accuracy between the Libre and Dexcom CGMs. Although 11 out of 15 studies reported slightly better outcomes for Dexcom, the combined mean difference was 0.96, with the 95% confidence interval crossing zero. This indicates equivalent accuracy between devices when compared to gold-standard glucose measurements. Conclusions: Libre performance can be influenced by factors such as hypoglycemic episodes, rapid glucose shifts, physical activity, sensor placement, and variability at the start or end of the sensor wear period. However, studies support its acceptable accuracy in diverse groups, including pediatric and pregnant populations. Further research is recommended on trend arrow reliability for clinical decisions.</p
Retinopathy Grading and Fundus Image Synthesis Using Generative Adversarial Networks (GANs)
No full textObjective: Recent advances have enabled the generation of synthetic retinopathy images using Generative Adversarial Networks (GANs). These models can produce high-resolution images that are nearly indistinguishable from real patient fundus photographs. Such synthetic images hold significant potential, particularly for medical education and algorithm development. This study outlines recent developments in GAN-based fundus image synthesis and evaluates their potential impact on diabetic patients and healthcare systems. Method: Multiple publicly available retinal fundus image datasets have facilitated research in this field, providing benchmarks for training and evaluating machine learning models. GANs offer a unique advantage by learning deep representations from limited data, reducing the need for expert annotations while preserving clinically relevant features in the images. This makes GANs particularly valuable for addressing data scarcity and class imbalance in retinopathy grading. Result: While GAN-based vessel segmentation methods have shown promise, they still struggle to detect extremely thin retinal vessels. Future improvements may involve incorporating prior anatomical knowledge to enhance segmentation accuracy. Some current GAN approaches can synthesize overall retinal structures but fail to replicate detailed vascular patterns. A potential solution is to generate synthetic vascular structures separately and integrate them into GAN-generated images. Conclusion: Since the emergence of deep learning, over a hundred DL-based methods have been proposed for retinopathy grading, primarily focused on vessel segmentation. However, the application of GANs—introduced around 2018— has shown superior performance in classification tasks due to their ability to generate high-quality synthetic fundus images. This capability addresses challenges such as class imbalance and data scarcity, making GANs a powerful tool in medical imaging. Continued research in this area could lead to substantial advancements in automated diagnosis and healthcare training.</p
Investigating Mesoscale, Submesoscale, and Mixing Ocean Dynamics in the Mozambique Channel Using Seismic and Simulation Datasets
No full textMesoscale and submesoscale structures and turbulent mixing play a vital role in transporting heat, carbon, and nutrients throughout the ocean, impacting ocean
circulation and climate. As the bridge between geostrophic balanced mesoscale eddies and smaller scale turbulent dissipation, submesoscale processes (10 m–10 km, e.g. fronts, filaments, vortices, and wakes) facilitate the forward energy cascade. A lack of observations with sufficient spatial and temporal resolution however hamper understanding interior ocean dynamics. Better capturing mesoscale–submesoscale coupling at complex topography is particularly important as such regions act as hotspots of ocean mixing. One promising method is the use of active acoustics or Seismic Oceanography. This approach can map whole swaths of ocean in several
hours with a resolution of O(10-100) m and penetrate to the seafloor. Here, a marine seismic dataset, collected in January-March 2016 has been to investigate the detail of subsurface flows in the Mozambique Channel. The Mozambique Channel is an ideal laboratory to study mesoscale-submesoscale interactions using seismics: here water masses from the Indian and Atlantic ocean collide and interact, driving sharp
thermohaline gradients and intense current flows interact with complex bathymetry to produce an energetic eddy field. First, a standard processing strategy was applied to seismic sections: frequency
filtering, seabed muting, direct wave removal, and f-k filtering, followed by velocity analysis, amplitude correction, stacking and deconvolution. Detailed images of phenomena including anticyclonic and cyclonic eddy distortions, submesoscale lenslike
structures, and water mass boundary are presented.
A combined inversion was developed and applied using integrated root mean square sound velocity analysis (VA) with iterative Markov Chain Monte Carlo
(MCMC) techniques. This approach successfully retrieved thermohaline fields with quantify uncertainties, without the requirement for coincident hydrographic data (ΔT ∼ 1.65 − 2.5◦C, ΔS ∼ 0.08 − 0.5PSU). The dominant error source is converting
root-mean-square velocity to interval velocity via the Dix equation. The spatial distributions of five seismic reflection sections capturing anticyclonic eddy
(AE), cyclonic eddy (CE), and non-eddy conditions were interpreted using inverted temperature-salinity fields. The AE enhanced downwelling, deepened Sub-Tropical Surface Water/South Indian Central Water watermasses to ∼700 m, and trapped Red Sea Water (RSW)/ Antarctic Intermediate Water (AAIW) as lenses at 800–1200 m. The CE drove upwelling, shoaling Sub-Tropical Surface Water (STSW)/ South Indian Central Water (SICW) and disrupted RSW. In non-eddy conditions, stratification remained more stable and North Atlantic Deep Water (NADW) was present.
These results highlight the role of eddies in shaping vertical water mass structure, stratification, and mixing.
Diapycnal diffusivity and turbulent dissipation rates, deduced from seismically derived horizontal spectra, were quantitatively investigated and verified. The Mozambique Channel was found to be a mixing hotspot with area mean log10(K) ∼ 0.5 − 1 an order of magnitude above open ocean values. Mixing intensity however differed with eddy polarity. Mixing was week in the AE core (K ∼ 10−6.5 m2 s−1) with localized
hotspots near boundaries and submesoscale features (∼ 10−4.2 m2 s−1). The CE exhibited stronger mixing (∼10−3.5 m2 s−1), especially near intense surface stirring and near the seafloor, linked to internal wave generation and bottom boundary layer instabilities. Non-eddy conditions also displayed bottom enhanced mixing (K ≈ 10−4 m2 s−1). Finally, simulations from the Coastal and Regional Ocean Community model
(CROCO) were combined with seismic data to analyze AE and CE dynamics. CROCO captured key mesoscale properties, water mass structure, and submesoscale features observed in seismic data. For example, a simulated AE showed deepened isopycnals
and lens-like submesoscale structures embedded within the eddy peripheries. One such cyclonic submesoscale structure was tracked and found to be generated by barotropic instabilities associated with AE boundary-slope interaction. Another anticyclonic lens-like structure was generated where the AE encountered the Davies Ridge. Energy conversion diagnostics indicate that mean-to-eddy kinetic energy
transfer sustain submesoscale coherent vortices in AEs, while weak or negative conversion in CEs suppresses them. Seismic structure observations support this
mechanism. In summary, results of this thesis reveal how eddy polarity and topography jointly regulate stratification, mixing, and interior submesoscale structure formation, advancing understanding of water mass transformation in the Mozambique Channel. Beyond the regional study, the approach establishes a framework for exploiting historical seismic data to investigate submesoscale processes more broadly.</p
Multimodal Prediction of Type 1 Diabetes
No full textType 1 diabetes often presents abruptly with life-threatening symptoms; however, its autoimmune pathogenesis typically unfolds over many years. After autoimmunity has been established, the rate of progression to clinical disease varies markedly between individuals, creating both an opportunity and a challenge. Early identification can reduce diabetic ketoacidosis at diagnosis and enable the use of preventative therapies; however, disease heterogeneity limits the accuracy and transportability of risk estimates used to guide screening, follow-up, and treatment decisions across settings and age groups. This thesis aims to investigate how demographic, immunologic, metabolic, and genetic factors jointly explain variation in progression from islet autoimmunity to clinical type 1 diabetes, and how this knowledge can inform screening and precision risk stratification.
Using longitudinal data from two major type 1 diabetes studies with distinct screening strategies, this work evaluates prediction model transportability, characterises age- and sex-related variation in disease progression, and refines metabolic staging in adults.
First, a multivariable risk prediction model combining genetic risk, autoantibody measures, age, and family history is developed and externally evaluated across cohorts. While key predictors are shown to be consistent, accurate risk estimation requires recalibration to the target screening context. This model is translated into a web-based risk calculator to support individualised risk communication and clinical or research decision-making. Next, sex differences in progression are examined, demonstrating that sex-related risk is not constant but varies with age and autoimmune status, with evidence of a developmental shift in progression trajectories around early adolescence. Differences between adult and paediatric presymptomatic type 1 diabetes are then investigated within the TrialNet cohort, revealing distinct immunologic and genetic profiles and slower average progression in adults, while showing convergence in progression rates once dysglycaemia and multiple autoantibodies are present. Finally, HbA1c is evaluated as a marker of metabolic progression in autoantibody-positive adults, demonstrating that age-related increases in HbA1c inflate dysglycaemia classification without necessarily reflecting increased risk of type 1 diabetes.
These findings collectively show that presymptomatic type 1 diabetes progression is shaped by interacting biological and demographic factors, and that identical biomarkers or thresholds do not carry uniform meaning across the life course. This thesis advances a framework for contextual interpretation of individual risk in type 1 diabetes screening and supports the development of more precise, equitable, and clinically meaningful early identification strategies. Future implementation will require broader external validation, increased ancestral diversity in study populations, and prospective evaluation of risk tools within real-world screening pathways.</p
Applied research in Parkinson’s disease and related conditions: mapping the funding and evidence landscape
No full textWhy this report mattersParkinson’s disease and related Parkinsonian conditions pose a growing challenge for health and care systems in the UK. These conditions are progressive and life-limiting. Treatments remain largely symptomatic, and there is still no proven disease-modifying therapy. People affected report unmet needs across diagnosis, treatment, care and support. This report was commissioned to inform policy and research commissioning by examining whether current research funding and evidence align with applied needs. It maps UK research funding since 2014 and international research evidence since 2019 across applied research categories of prevention, diagnosis, treatment and care, with health inequalities considered throughout.What we didWe analysed the 670 UK-based research grants awarded since 2014, representing around £280 million in investment. We mapped 9,196 applied research papers and systematic reviews published internationally between 2019 and 2024. Funding and evidence were categorised into prevention, diagnosis, treatment and care.What we foundApplied research is relatively under-represented. Nearly two-thirds of UK Parkinson’s research funding supports disease mechanisms and pre-clinical research. Only around one-quarter is directed towards applied research.Funding within applied research is uneven. Treatment dominates, accounting for around two-thirds of applied research grants and nearly a quarter of applied research funding. Diagnosis receives a smaller but substantial share. Care attracts limited funding, and primary prevention is almost entirely absent, receiving just 1.1% of total applied research investment.The research evidence broadly reflects the same patterns. Most applied studies focus on treatment, particularly drug therapies and motor symptoms. There is far less evidence on prevention and care, and limited attention to common symptoms such as anxiety, depression, cognitive impairment, fatigue, pain and communication.Prevention is a critical gap. Although many studies examine factors associated with Parkinson’s risk, almost none test approaches or interventions designed to reduce risk or delay onset. The relatively high number of reviews in this area reflects growing interest and suggests there may be potential to increase capacity for prevention research.Care research is under-developed. Much of the care literature is descriptive, with very few intervention studies, limiting the evidence available to support service development and delivery. Palliative and end-of-life care are particularly under-researched. Inequalities and capacity issues persist. Very few grants or studies explicitly consider health inequalities beyond age and sex, limiting insight into differential needs and outcomes. Funding is also geographically concentrated, with over half awarded to institutions in London and Oxbridge, raising questions about regional research capacity and equity.What this means for policy and fundersOverall, there is a mismatch between applied research needs and the current portfolio. There is potential to increase the focus on prevention and care, strengthen research on non-motor symptoms and lived experience, embed inequalities as a core design consideration, and support more geographically and disciplinarily diverse research teams. A more coordinated, person-centred applied research strategy, developed with involvement of people with lived experience, has the potential to deliver greater value from future investment and improve outcomes for people affected by Parkinson’s and related conditions.</p
Advancements in model predictive control for fully automated glucose regulation in type 1 diabetes
No full textBackground and Aims: Model Predictive Control (MPC) is emerging as a leading strategy for automating glucose management in individuals with Type 1 Diabetes, particularly within fully closed-loop (FCL) systems. This review assesses the clinical performance of current FCL technologies and explores future directions for improving their efficacy through the integration of advanced algorithms. Methods: A detailed literature review was performed using databases including PubMed, TRIP Pro, and Web of Science. Search terms such as “Fully Closed Loop,” “Type 1 Diabetes,” “Artificial Pancreas,” “Continuous Glucose Monitoring,” and “Model Predictive Control” were used. Boolean operators “AND” and “OR” refined the search, allowing for a comprehensive comparison of various FCL approaches and their technological components. Results: Recent evidence shows that MPC-based FCL systems outperform hybrid closed-loop (HCL) setups using Proportional-Integral-Derivative (PID) algorithms. MPC models achieved higher time-in-range (TIR) outcomes (74.4% vs.63.7%, P = 0.020) and superior post-meal glucose control. Despite these improvements, many systems still struggle to maintain TIR consistently above the clinical target of 70%. Persistent challenges include managing postprandial hyperglycemia, insulin delivery delays, and device limitations. Emerging solutions include nonlinear MPC (NMPC) for dual-hormone delivery, adaptive reinforcement learning via k-Policy Iteration, and pulse-modulated control systems mimicking natural insulin secretion. Conclusions: Although promising in simulations, these innovations require further clinical testing. Key barriers include sensor inaccuracies, glucagon formulation issues, cost, and user variability. Future research should prioritize real-world trials that integrate adaptive algorithms, dual-hormone systems, and account for lifestyle factors like exercise and meal variability.</p
Assistive technology for scotoma with diabetic retinopathy
No full textObjective: This research presents the development of novel computer-based simulations of scotoma (blind spots) caused by diabetic retinopathy. These simulations investigate impacts of scotoma on reading and activities of daily living. The study explores assistive technologies aimed at mitigating the functional limitations with scotoma. Key components include: (i) simulation of visual blind spots, (ii) performance of search tasks with simulated scotoma, (iii) reading with artificial blind spots, (iv) use of virtual reality headsets to simulate scotoma, (v) development of assistive visual technologies, and (vi) scotoma and real-life implications. Method: An assistive device was developed to remap visual information from areas of the retina affected by scotoma onto peripheral retinal regions. This visual remapping system aims to restore functional vision, projecting obscured content into visible areas. Evaluation of the system indicates its potential as a promising assistive technology for individuals with central vision loss due to scotoma. Result: Simulated scotoma scenarios were deployed across real-world environments on university campus, including reading text displays, navigating classrooms, interpreting directional signage, and participating in physical activities. In each setting, visual content occluded by the scotoma was algorithmically remapped to adjacent healthy peripheral vision zones. These simulations were applied to still images and video recordings. The system was demonstrated to university researchers and NHS-affiliated ophthalmologists, who provided positive feedback on utility and realism. Conclusion: The visual remapping system received favorable responses from academic and clinical experts. This also identified limitations in current simulated scotoma models, in terms of realism and accuracy. Future work should enhance fidelity of simulations, incorporating direct feedback from individuals with scotoma, ensuring simulated experiences closely reflect real-world visual impairments.</p
Synthetic retinal image generation and disease classification with GAN models
No full textBackground and Aims: Generative Adversarial Networks(GANs) have recently emerged as powerful tools for creating synthetic retinal images that closely resemble real fundus photo-graphs. These realistic simulations have potential applications in clinical training, diagnostic support, and medical AI development. This study highlights current progress in GAN-driven retinal image synthesis and discusses its potential impact on diabetic retinopathy (DR) management, patient care, and health-care system efficiency. Methods: Several open-access databases containing retinal fundus images have fueled progress in this field, serving as essential resources for training and validation. GANs are particularly suited for learning detailed image features even from relatively small datasets, helping mitigate the need for extensive expert labeling. This makes them especially useful in situations where there is a shortage of annotated data or significant class imbalance in disease severity levels. GANs preserve crucial clinical details while generating synthetic data, supporting both automated grading and enhanced data-set augmentation. Results: Although GAN-based vessel segmentation has achieved notable advances, accurately capturing the finest vascular details remains challenging. Current systems can replicate general retinal anatomy but often fall short in reproducing intricate microvascular structures. One potential strategy is to first create separate synthetic vessel maps and then embed these into the generated fundus images for more anatomically realistic outputs. Conclusions: Since their introduction into ophthalmology, GANs have expanded the possibilities for automated retinopathy analysis. Their ability to generate high-fidelity synthetic data helps overcome data limitations, paving the way for improved diagnostic models, advanced educational tools, and more balanced training datasets in medical imaging.</p
Evolutionary neural networks for blood glucose prediction in diabetes
No full textObjective: This study introduces a novel evolving neural network approach for forecasting blood glucose levels in individuals with diabetes. The proposed method demonstrates superior predictive accuracy compared to a traditional backpropagation neural network. Maintaining stable blood glucose levels is essential for individuals with diabetes, as prolonged hyperglycemia can lead to severe health complications. Method: The research employs an evolutionary computation technique to evolve a neural network architecture using neuro- evolution. The optimized neural network is then applied to predict changes in blood glucose levels. To evaluate performance, the evolved network was compared against a standard backpropagation neural network using continuous glucose monitoring (CGM) data. Results showed that the evolved model consistently outperformed the conventional model in prediction accuracy. Result: The evolved neural network achieved higher accuracy than the standard backpropagation model. The use of a genetic algorithm yielded promising results for predicting CGM data. Data was partitioned into training, validation, and testing sets, enabling effective model development, comparison, and evaluation. Conclusion: This approach offers a valuable tool for individuals with diabetes, providing advance warnings of potential blood glucose fluctuations. Early prediction can help prevent episodes of hypoglycemia and hyperglycemia, thereby reducing the risk of complications and lowering healthcare costs. Future work should explore the use of larger and more diverse CGM datasets, as well as assess the impact of different data recording intervals—such as 1, 5, and 15 minutes—on prediction performance.</p
Castles and Communities: Using the taskscape to explore the attitudes, activities, and lived experience of wider communities in castle landscapes between 1050 - 1550.
No full textThis thesis utilises the concept of the taskscape to explore the impact of the medieval castle in England on lived experiences in the landscapes surrounding them between 1050 to 1550. Recent shifts in the focus of castle studies have demonstrated the need to focus on castles as sites of everyday activities, inclusive of a range of social experiences beyond those of the elite occupants. These insights will be gained through the use of the taskscape, viewing the medieval world as a series of mutually attentive interlocking temporalities, helping contextualise the castle within ongoing dwelling.
In order to bridge the divide between theory and methodology, a integrative methodology is suggested, using geographical information systems to approach the sensory and temporal dimensions of tasks. The form of the landscape is modelled using a retrogressive analysis of historic maps and LiDAR imagery. Portable Antiquity Scheme data, excavations, and historical and archaeologically attested sites and places are then integrated into a model of the physical and social topography of the medieval world. Visibility and movement analysis then informs analysis of the motion and sensory experiences inherent to the performance of dwelling tasks.
This methodology is applied to three pairs of case studies, helping focus on different dimensions of the taskscape, whilst also reflecting the variations in castle and landscape form across dispersed, nucleated, rural, and urban settings. Ultimately, three conclusions are reached. Firstly, everyday movements of materials and individuals around the castle ensured that it maintained a central place within the lived experience of non-elite communities, and consequently in their production of identity. Secondly the castle adopted long term temporalities of memory, geology, and natural rhythms into a display of social hierarchies and parallel temporalities. Finally, the impact of the castle was relative to the strength, intensity, and momentum of surrounding tasks.</p