Real-world movement assessments enable continuous monitoring of gait impairments in people with Parkinson’s disease (PD), but their interpretation is challenging due to fluctuations arising from the context-dependent nature of daily-life walking. This study analysed real-world gait outcomes and contextual data (medication intake, weather, and location) collected from twenty people with PD. Machine learning regression models were used to quantify the contribution of contextual factors to variability in gait outcomes. Contextual factors explained 30–40% of the variability in real-world gait outcomes, with the highest predictive performance observed for gait speed (R² = 0.39) and the lowest for cadence (R² = 0.30). Feature importance analysis indicated that bout-specific characteristics, start time of the bout, and time since medication intake were the primary predictors. These findings demonstrate that real-world gait in people with PD is substantially influenced by contextual factors, highlighting the need for context-aware analysis when interpreting gait outcomes.

Introduction Conventional clinical assessments do not fully capture how Parkinson's disease (PD) affects mobility in daily life. Integrating digital mobility outcomes (DMOs) from wearable devices with GPS-derived contextual data could provide richer insight into real-world mobility, yet this approach remains largely unexplored. Similarly, data-driven modeling of DMO distributions, such as walking speed, may reveal clinically relevant changes in mobility that are obscured by averaged measures. This study (i) examined how indoor-outdoor context enhances interpretation of real-world mobility, and (ii) applied Gaussian Mixture Modeling (GMM) to characterize data-driven patterns within walking speed distributions in people with PD.Methods Fifty-two people with PD (PwP) and 19 older adult controls were recruited from the CiC and Mobilise-D studies. DMOs were estimated from a single wearable device, and indoor-outdoor location was synchronized with GPS data from a smartphone. GMM was applied to estimate the optimal number of walking speed modes. Generalized linear models compared DMOs between indoor and outdoor contexts and between cohorts, adjusting for age and sex.Results Thirty-nine PwP and 17 controls had valid contextual data. Both cohorts performed significantly more indoor than outdoor walking bouts, with longer walking durations outdoors. Only controls walked significantly slower and with shorter strides indoors versus outdoors, while both groups showed longer stride duration indoors. Between-cohort differences emerged only outdoors, with PwP exhibiting higher cadence. Most participants across both cohorts displayed three walking speed modes, which were associated with medication dosage and motor severity.Discussion This study demonstrates the potential of GPS-derived contextual information to enhance interpretation of real-world mobility outcomes in PD. Walking speed modes show promise for capturing novel clinical insight, though further technical and clinical validation is required to establish their robustness and clinical relevance.

Digital mobility outcomes (DMOs) derived from wearable sensors characterise mobility in daily life and offer a promising means of monitoring disease progression. Yet most DMO studies examine one disease at one visit; they do not model how multivariate DMO relationships with multiple clinical outcomes evolve jointly across diseases. Technically, existing temporal multi-task frameworks can model progression within an individual disease, but they do not jointly model multiple prediction outcomes across diseases, particularly when disease cohorts do not share participants. To address these gaps, we propose DeMMO, an interpretable framework for longitudinal, multi-disease, and multi-outcome learning. DeMMO represents each disease-outcome objective by a longitudinal DMO coefficient matrix and combines temporal regularisation with stable and visit-specific feature selection. Its central technical contribution is an automatic cross-disease and cross-outcome relation-learning mechanism that learns signed relations directly from these longitudinal mappings, enabling selective information sharing without paired participants. We evaluate DeMMO on the recently released, large-scale, multicentre Mobilise-D dataset, which provides a new opportunity to study 24 harmonised real-world DMOs over five visits across multiple mobility-limiting conditions. Against nine strong linear, longitudinal, and deep-regression baselines, DeMMO achieves the best overall and outcome-specific prediction performance, with significant improvements over the strongest baselines. Stability selection further identifies reliable longitudinal DMO patterns for subsequent clinical validation and disease monitoring. The implementation code and experimental results are available at https://github.com/menghui-zhou/DeMMO.

The increase in privacy concerns and the introduction of privacy and data protection legislation compel organisations to reevaluate their practices regarding traditional machine learning. The aggregation and management of users’ private data on the central server may contravene regulations if not properly administered. Federated learning provides a technique that eliminates the necessity of uploading users’ data to the server. It facilitates substantial learning by collaboratively training on each client’s devices and pooling the model gradient changes. Federated learning, augmented with a proxy as an intermediary and encrypted model parameters, will enhance anonymity, privacy, and data protection against malicious threats, including membership inference adversaries. Nonetheless, encrypted data incurs costs for customers’ communication and data size that exceed twice the original size. Our paper seeks to resolve these issues. We present two secure approaches for effective communication in an anonymous encrypted federated learning framework as our contribution. Additionally, our experiments demonstrated that it is feasible to attain equivalent communication costs as in non-encrypted scenarios. We provide recommendations in the conclusion for the effective implementation of privacy-preserving federated learning in the area of personal devices.

Almost all societal grand challenges, whether concerning the environment, health, well-being, or the development of sustainable economic models, have at their heart a need to understand people’s behaviour. However, uniting data and insights across disparate fields requires an explicit and shared understanding of concepts, variables, and ideas (e.g., how to characterise and differentiate behaviours). Ontologies provide a mechanism for creating this explicit and shared understanding and are starting to be developed and used in the social and behavioural sciences. This paper proposes an online co-design approach to use and develop ontologies of behaviour to specify the characteristics of behaviour (e.g., habitual, changeable, effortless) and studies that investigate behaviour as part of a project designed to understand how behaviours are related. We report on our experience of collaborative co-development of ontologies using real-time interactive tools and reflect on the benefits and challenges of our approach. We also offer a set of recommendations for researchers interested in applying such methods to co-develop ontologies. The work contributes to efforts to understand the characteristics of behaviour and enable these to be used to understand questions about behaviour (e.g., is poor sleep associated with greater engagement in habitual behaviours?).

Privacy concerns have escalated due to companies' misuse of user data and the occurrence of data breaches and leaks worldwide. Uploading personal data from personal devices to a central server over the network poses a danger in obtaining an inference. Hence, a different approach is needed for this scenario. Federated learning (FL) enables collaborative training on devices while maintaining the privacy of user data. FL originally aimed to address privacy concerns but is vulnerable to certain privacy attacks. Although certain privacy-enhancing strategies are available, researchers are actively seeking a more effective option. This research suggests two privacy improvement methods using proxies as a better option for personal devices in a FL environment, achieving good performance and cost effective without accuracy loss. We studied and assessed how the methodology compared to other methodologies. Finally, we discussed how this proposed technique can address the limitations of other techniques and possible collaborations with them.

Background: Young people with life-limiting conditions have unmet psycho-spiritual needs. Dignity Therapy is a psycho-therapeutic end-of-life intervention that addresses life review, meaning making and legacy leaving. Studies of Dignity Therapy including young people are limited. Aim: To co-design a digital Dignity Therapy-based intervention for young people with life-limiting conditions (DIGNISPACE). Design: A qualitative study including focus groups (n=5) with hospice-based healthcare professionals (n=23), semi-structured interviews with young people with life-limiting conditions (n=13) and family carers (n=12). Data were analyzed using framework analysis. Results: Three main themes were derived; intervention purpose, amendments to the Dignity Therapy question protocol and content for a digital application. Findings influenced the development of DIGNISPACE and a model of dignity in young people with life-limiting conditions. Conclusions: DIGNISPACE addresses the psycho-spiritual needs of young people with life-limiting conditions, including tenets of the model of dignity in young people with life-limiting conditions in its content and delivery.

With the rise of privacy-preserving machine learning, browser-based federated learning (WebFL) offers an accessible way to train models without exposing sensitive data. Hardware acceleration via Graphics Processing Units (GPUs) has been proposed as a means to reduce training latency and resource consumption, yet the effects of browser-level GPU configuration flags remain unclear. This study systematically evaluates the impact of 43 GPU-related flags in Chrome and Firefox on anonymous federated learning performance. The results show that GPU flag tuning does not deliver meaningful improvements in training duration, transformation time, or communication cost. The only notable outlier was Firefox's webgl. angle. force-warp, which caused severe slowdowns due to CPU-based emulation. In Chrome, certain backends yielded modest memory efficiency gains ($\leq \mathbf{2 0 \%}$), but overall behaviour remained stable. The novelty of this work lies in documenting a negative but important result: GPU flag optimisation at the browser level offers limited practical value for accelerating WebFl. These findings provide actionable guidance-practitioners should rely on default stable configurations, while researchers should direct optimisation efforts toward communication, model design, and memory management rather than browser GPU toggles.

Passenger Name Record (PNR) data is essential for transportation analysis and security research, particularly in surveillance and threat detection. However, stringent security and privacy concerns limit access to real PNR data. This study presents a methodology for generating synthetic PNR data that not only replicates statistical properties but also reconstructs passenger social networks, models travel behaviours and preserves individual travel histories for security and mobility analysis. Our approach generates detailed, individual-level data—including passengers, bookings, and flights—while maintaining spatial, temporal, and chronological consistency to ensure realistic movement patterns while upholding privacy. This methodology offers a privacy-preserving alternative for transportation security and behavioural research, expanding access to high-quality data for future studies.

Quantifying task similarity is crucial in continual learning, enabling models to better mitigate catastrophic forgetting and facilitate knowledge transfer across tasks. However, existing similarity measures often demand extensive data and computation resources, resulting in low efficiency and limiting practical applications. To address this challenge, we introduce SPOT(Single-batch Probe Of Task-similarity), a novel task similarity measure that requires only a single batch of data to quickly estimate task similarity before training on a new task. SPOT leverages the change in the empirical loss between old and new tasks to quantify task relationships, offering a low-cost, efficient solution for task similarity estimation in continual learning. Experiments on three public datasets and one real-world dataset show that 1) task similarity and forgetting are negatively correlated. 2) SPOT can efficiently predict the forgetting risk with one batch of new task data. 3) Forgetting is most severe for tasks with significant semantic distinctions. Our findings indicate that SPOT serves as a passive yet efficient tool to predict catastrophic forgetting risk before training, facilitating continual learning with minimal computational overhead. This research provides new insights into task similarity quantification and has strong potential for deployment in resource-constrained environments. Code is available at https://github.com/wangxulong/SPOT.

Background Ontologies are frameworks for representing information that promote clarity, consistency and coherence, reduce the fragmentation of knowledge, and allow datasets and knowledge to be linked across studies, disciplines and domains. To enable this, it is important to identify how concepts of interest (‘classes’) are represented in different ontologies and evaluate the extent to which such classes align (i.e., are ‘interoperable’). This study aims to provide a method for doing this. Methods An automated tool using Meta’s Llama 3 language model was developed and used to compare artificial intelligence (AI) and human approaches to matching ontology classes. The automated tool was then integrated into a hybrid method for identifying classes that appear to refer to the same thing across pairs of ontologies. The method was evaluated by three behavioural scientists who used it to identify similar classes in two ontologies and provided feedback on their experience. Results The automated tool identified a larger number of potential matches than human-led review, so was used to generate a shortlist. The evaluation of the method produced mixed results. Users agreed which classes were identical or essentially the same across contexts, but none of the users identified similar classes that could be imported into an ontology without causing a contradiction or conflict. Users typically found using the method difficult, but many of the challenges related to using ontologies, rather than to the method specifically. Conclusions A combination of automated and human processes appears to be a feasible way to assess the interoperability of ontology classes. While further refinement is needed along with tools and resources that enable the use of ontologies by a broad range of researchers, the study provides a workable method for matching ontology classes in the behavioural and social sciences and offers a practical guide to support its implementation.

On-device learning, such as federated learning, is gaining more popularity. It benefits users with faster inference and privacy preservation. However, the heterogeneity of personal devices makes its deployment not easy. With the increasing need for an on-demand learning platform, the web browser has become one of the leading solutions due to its availability and interoperability. Nevertheless, there is still a lack of research on evaluating the behaviour of federated learning on web browser platforms. This includes evaluating their compatibility, convergence in inference results, and performance. Our paper tries to address these concerns. Throughout our experiments, we found that there are still inconsistencies in inference results, compatibility issues, and varied performance among these platforms. Besides experiment analysis on this subject, we also recommend a model-platform-device compatibility report as our contribution.

Alzheimer's disease (AD) is the leading cause of dementia worldwide, characterized by its gradual progression and the subtle variations across disease stages, which pose significant challenges for accurate diagnosis. While deep representation learning algorithms have shown promise in the early detection of AD using MRI data, existing approaches often overlook the meaningful relationships between continuous labels in AD progression, and the learned representations frequently lack interpretability due to the black-box nature of deep learning models. To address these limitations, we propose ICReL (Interpretable Continuous Representation Learning), a novel concept based on coding rate principles that captures continuous representations across AD stages while maintaining a high degree of interpretability. Extensive experimental results on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset demonstrate that ICReL not only outperforms multiple baseline methods using 2D slice or 3D MRI in terms of learning continuous representations, but also exhibits enhanced robustness to label corruption and superior predictive performance. This work offers a new, interpretable approach to representation learning for computer-aided diagnosis of neurodegenerative diseases.

Investigations of children experiencing inflicted injuries is often initiated once admitted into the emergency department for injuries, and involves understanding the complex interactions between a range of different factors such as health conditions, fracture biomechanics, family history, carer accounts and so on. In this position paper, we propose the use of ontologies for capturing case details of such radiology investigations in order to create an initial knowledge base of retrospective cases, to be further expanded in the future. We discuss how we developed our ELECTRICA (ELEctronic knowledge base for Clinicians, Trainers and Researchers in Child Abuse) ontology and the different components of the ontology. We are currently using this ontology to create a knowledge base, to be used as a vision demonstrator to access larger datasets. In the longer term, we would like to use the knowledge base to support clinicians and radiologists in making decisions on current cases by offering a mechanism for searching historical cases, identifying similar cases and offering insights on risk factors that may have been missed during investigations.

This study explores the correlation between residents’ subjective assessments of urban neighbourhoods, obtained through virtual walkthroughs, and objective measures of deprivation. Our study was set within a specific city in the United Kingdom, with neighbourhoods selected based on Indices of Multiple Deprivation (IMD). We invited residents in the UK through Prolific, a crowdsourcing platform. Employing complete case analysis, TF-IDF keyword extraction, the Kruskal–Wallis test, and Spearman’s rank-order correlation, our study examines the alignment between subjective assessments and existing deprivation measures (IMD). The results reveal a nuanced relationship, suggesting potential subjective biases influencing residents’ perceptions. Despite these complexities, the study highlights the value of virtual walkthroughs in offering a holistic overview of neighbourhoods. While acknowledging the limitations posed by subjective biases, we argue that virtual walkthroughs provide insights into residents’ experiences that potentially complement traditional objective measures of deprivation. By capturing the intricacies of residents’ perceptions, virtual walkthroughs contribute to a more comprehensive understanding of neighbourhood deprivation. This research informs future endeavours to integrate subjective assessments with objective measures for robust neighbourhood evaluations.

Machine learning techniques for predicting Alzheimer's disease (AD) progression can substantially help researchers and clinicians establish strong AD preventive and treatment strategies. However, current research on AD prediction algorithms encounters challenges with monotonic data form, small dataset and scarcity of time-continuous data. To address all three of these problems at once, we propose a novel machine learning approach that implements the 4D tensor multi-task continual learning algorithm to predict AD progression by quantifying multi-dimensional information on brain structural variation and knowledge sharing between patients. To meet real-world application scenarios, the method can integrate knowledge from all available data as patient data increases to continuously update and optimise prediction results. To evaluate the performance of the proposed approach, we conducted extensive experiments utilising data from the Alzheimer's Disease Neuroimaging Initiative (ADNI). The results demonstrate that the proposed approach has superior accuracy and stability in predicting various cognitive scores of AD progression compared to single-task learning, benchmarks and state-of-the-art multi-task regression methods. The proposed approach identifies structural brain variations in patients and utilises it to accurately predict and diagnose AD progression from magnetic resonance imaging (MRI) data alone, and the performance of the model improves as the MRI data increases.

Alzheimer’s disease (AD) is the leading cause of dementia worldwide, characterized by its gradual progression and the subtle variations across disease stages, which pose significant challenges for accurate diagnosis. While deep representation learning algorithms have shown promise in the early detection of AD using MRI data, existing approaches often overlook the meaningful relationships between continuous labels in AD progression, and the learned representations frequently lack interpretability due to the black-box nature of deep learning models. To address these limitations, we propose ICReL (Interpretable Continuous Representation Learning), a novel concept based on coding rate principles that captures continuous representations across AD stages while maintaining a high degree of interpretability. Extensive experimental results on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset demonstrate that ICReL not only outperforms multiple baseline methods using 2D slice or 3D MRI in terms of learning continuous representations, but also exhibits enhanced robustness to label corruption and superior predictive performance. This work offers a new, interpretable approach to representation learning for computer-aided diagnosis of neurodegenerative diseases.

Mobility-related ambulatory activity (e.g., walking) is essential to healthy cognitive and functional ageing. Wearable technology (e.g., accelerometers/inertial measurement units; IMUs) allows us to objectively and continuously capture digital mobility outcomes (DMOs), e.g. volume, pattern and variability of walking activities. Continuous digital mobility assessment has been shown as feasible and acceptable to older adults with and without dementia. However, interpretation of DMOs is limited as we do not yet understand the impact of different social and environmental contexts on walking activity. For example, DMOs may be influenced by the volume of walking that one’s partner participates in, or by the walkability of their local area. Understanding the impact of social and environmental contexts on DMOs will support development of socio-ecological strategies to support mobility in ageing. We aim to conduct a feasibility study in older adults with the following objectives: (1). Objectively assess DMOs in older adult dyads using digital mobility tools (i.e., IMUs/GPS); (2). Identify key social and environmental influences on DMOs through mixed-methods exploration; (3). Develop a novel analytical approach combining DMOs with GPS data to assess independence/interdependence in dyads’ walking activities. We will recruit 20 older adult dyads in North-East England for an observational cross-sectional study. Walking activities will be recorded continuously for seven days using an IMU attached to participants’ lower backs. DMOs include volume (e.g., daily steps), pattern (e.g. mean bout length) and variability (of bout length) of walking activities. Simultaneously, participants will carry a GPS device (i.e., smartphone) to monitor excursions outside the home. Questionnaires will capture information on cognition, function, falls risk, exercise motivation, wellbeing, relationship mutuality, spatial navigation, and walkability of local area. Participants will complete a “mobility diary” for the assessment period (e.g., daily journeys, motivations/perceptions/familiarity of journeys). Preliminary results will be presented regarding associations between social/environmental contexts with DMOs via flexible Bayesian statistical models and thematic qualitative analysis. This study will assess the feasibility of the protocol and analysis strategies, with intentions of developing further research to examine the impact of social and environmental influences on DMOs in people with dementia and their carers.