69238 research outputs found
Sort by
Personalised Glucose Control through Forecasting and Reinforcement Learning
No full textCurrent time series models and reinforcement learning techniques for glucose management
in people with diabetes emphasise forecasting for early intervention and automating medication
infusions. Often, these approaches operate under partial observability with a sensor
measurement estimate of the state where the overall system dynamics are unknown
and unobserved. This thesis aims to develop an accurate, automated, personalised glucose
management system by leveraging personalised synthetic data through ordinary differential
equations to estimate unknown dynamics. We propose two techniques: for forecasting,
we continually tune the parameters of the ordinary differential equations during training to
minimise the root mean squared error of the synthetic glucose and the real glucose values
while incorporating the generated states as exogenous variables. For reinforcement learning,
we incorporate a forecasting component to predict future actions for user execution,
adding synthetically generated states to the forecasting and reinforcement learning observation
space. We evaluate our approach on in vivo datasets to reduce forecasting error scores
and to regulate blood glucose using reinforcement learning through action prediction. Our
experiments show a 20% average reduction in forecasting error scores and increased time in
range by 2% against real participant baselines
Node Classification on Graph Data with Global Learning
No full textNode classification is a core task in graph-based machine learning, where the goal is to predict labels or categories for nodes in a graph, leveraging the relational structure and attributes of the data. To enhance the expressiveness of node representations, various Graph Neural Networks (GNNs) have been proposed to aggregate information from neighbouring nodes, conducting message-passing within local receptive fields, a process also referred to as local learning. However, the suboptimal nature of the available graph structure, often characterized by noisy edges or missing edges, generally negatively affects the performance of node classification. To address this issue, this dissertation explores node classification by considering relations among all nodes, referred to as global learning, and further investigates opportunities and challenges for its application across different scenarios.
-We introduce Chain of Propagation Prompting (CPP) to enhance the expressiveness of node representations while reducing the dependency on label information. CPP involves designing a simple message-passing pattern, which we incorporate into node representations using graph contrastive learning. This simple pattern prompts multi-head self-attention-based layers to globally capture more complex patterns while minimizing their reliance on label information. Additionally, we implement majority voting to enhance the predictive confidence of multiple heads.
-We introduce Robust Node Classification under Graph and Label Noise (RNCGLN) to improve the robustness of node classification when both graph and label noise are present. By integrating local graph learning and global graph learning, RNCGLN can provide comprehensive information to enhance node classification performance. Additionally, we develop graph and label self-improvement modules to improve and supplement the quality of supervisory information. Consequently, RNCGLN leverages self-training and pseudo-label techniques to facilitate two self-improvement processes in an end-to-end learning framework.
-We introduce a Flexible-pass Filter-based Graph Transformer (FFGT) to resist adversarial attacks on graph data. Leveraging self-attention's ability to capture arbitrary graph filters, our self-attention layers with three heads capture multi-frequency representations across low-frequency, hybrid-frequency, and high-frequency ranges. Additionally, we designed graph learning and fusion modules to improve self-attention effectiveness in capturing designed information, yielding a flexible-frequency representation. Consequently, FFGT shows consistent resistance to adversarial perturbation in multiple datasets and against diverse adversarial attacks.
We conducted theoretical analyses and numerical evaluations of our proposed methods using diverse graph data. The experimental results show that our methods, leveraging global learning strategies, consistently outperform traditional Graph Neural Networks (GNNs) based on local learning. This superior performance is demonstrated across various scenarios, including diverse graph datasets, graph noise, label noise, and multiple adversarial attacks. Additionally, our theoretical analysis validates the effectiveness of each proposed method in addressing these challenges. Together, these findings confirm that our methods enhance expressiveness, improve robustness to noise, and strengthen resilience against adversarial attacks
Drug-resistant gram-negative bacterial infections in children in the Oceania region: review of the epidemiology, antimicrobial availability, treatment, clinical trial and pharmacokinetic data, and key evidence gaps
No full textGram-negative bacterial infections remain a major cause of morbidity and mortality in children and neonates globally, compounded by the rise of antimicrobial resistance. Barriers to paediatric antibiotic licencing lead to reduced availability of potentially effective agents for treatment. For children and neonates in the Oceania region, specific challenges remain including a paucity of surveillance data on local rates of antimicrobial resistance, and lack of availability of newer, more costly agents. In this review, we summarise available regional epidemiological data on the WHO priority pathogens: extended spectrum B-lactamase (ESBL)-producing Enterobacterales, carbapenem-resistant Enterobacterales (CRE), carbapenem-resistant Pseudomonas aeruginosa, and carbapenem-resistant Acinetobacter baumannii (CRAB). Paediatric clinical trial and pharmakinetic data for the antimicrobials recommended for treatment of these pathogens are reviewed, and paediatric knowledge gaps identified to inform future collaborative research
Imputation of missing clock times - application to procalcitonin concentration time course after birth
No full textThe time course of biomarkers (e.g., acute phase proteins) are typically described using days relative to events of interest, such as surgery or birth, without specifying the sample time. This limits their use as they may change rapidly during a single day. We investigated strategies to impute missing clock times, using procalcitonin for population modelling as the motivating example. 1275 procalcitonin concentrations from 282 neonates were available with dates but not sample times (Scenario 0). Missing clock times were imputed using a random uniform distribution under three scenarios: (1) minimum sampling intervals (8-12 h); (2) procalcitonin concentrations increase for postnatal days 0-1 then decrease; (3) standard sampling practice at the study hospital. Unique datasets (n = 100) were created with scenario-specific imputed clock times. Procalcitonin was modelled for each scenario using the same non-linear mixed effects model using NONMEM. Scenarios were evaluated by the NONMEM objective function value compared to Scenario 0 (∆OFV) and with visual predictive checks. Scenario 3, based on standard sampling practice at the study hospital, was the best imputation procedure with an improved objective function value compared to Scenario 0 (∆OFV: -62.6). Scenario 3 showed a shorter lag time between the birth event and the procalcitonin concentration increase (average: 12.0 h, 95% interval: 9.7 to 14.3 h) compared to other scenarios (averages: 15.3 to 18.7 h). A methodology for selecting imputation strategies for clock times was developed. This may be applied to other problems where clock times are missing
Variational inference for correlated gravitational wave detector network noise
No full textGravitational wave detectors like the Einstein Telescope and LISA generate long multivariate time series, which pose significant challenges in spectral density estimation due to the large number of observations as well as the presence of correlated noise. Addressing both issues is crucial for accurately interpreting the signals detected by these instruments. This paper presents an application of a variational inference spectral density estimation method specifically tailored for dealing with correlated noise in the data. It is flexible in that it does not rely on any specific parametric form for the multivariate spectral density. We provide the multivariate Whittle likelihood in a form that is easy to evaluate as it depends on a low-dimensional covariance matrix. To deal with very long time series, the method employs a blocked Whittle likelihood approximation for stationary time series. It utilizes the Cholesky decomposition of the inverse spectral density matrix to ensure a positive definite estimator. A discounted regularized horseshoe prior is applied to the spline coefficients of each Cholesky factor, and the posterior distribution is computed using a stochastic gradient variational Bayes approach. This method is particularly effective in addressing correlated noise, a significant challenge in the analysis of multivariate data from colocated detectors. The method is demonstrated by analyzing 2000 s of simulated Einstein Telescope noise, which shows its ability to produce accurate spectral density estimates and quantify coherence between time series components. This makes it a powerful tool for analyzing correlated noise in gravitational wave data
The Ecology of Academic Misconduct: The Role of Moral Attitudes, Disengagement and Integrity Climate in Academic Dishonesty
No full textThis study investigates the relationships among integrity climate, peer norms, and academic dishonesty, with a particular emphasis on moral disengagement. Drawing on data from the Research on Academic Integrity in New Zealand (RAINZ) Project, the study applies Structural Equation Modeling to examine how integrity culture and perceived peer cheating influence academic misconduct through moral judgment and moral disengagement. Adopting an ecological psychology perspective and guided by social cognitive theory, the study highlights the role of moral judgment as a mediator and moral disengagement as a facilitator. A series of competing models are tested to evaluate alternative pathways and confirm the robustness of the proposed theoretical framework. Results show that integrity climate is strongly associated with ethical student behavior and demonstrate the value of model comparison in clarifying the mechanisms of academic misconduct. Practical recommendations for promoting academic integrity are offered, with implications for educational practice as well as future research
Characterising the Genetic Heterogeneity of Motor Neuron Disease in New Zealand
No full textMotor neuron disease (MND) is a group of neurodegenerative diseases which are characterized by adult-onset progressive degeneration of motor neurons, of which amyotrophic lateral sclerosis (ALS) is the most common form. MND has a complex aetiology with more than 40 genes discovered that are either directly causative of, or are associated with it, yet despite these identified causal genetic associations, only 5-10% of people with ALS have a Mendelian family history of disease (familial). Rather, the majority of cases appear in families where seemingly only a single individual is affected (sporadic) with disease. New Zealand has among the highest mortality and incidence rates in the world, yet to date no genetic studies have been undertaken in New Zealand to determine whether there is a genetic basis for this high rate of MND.
This thesis highlights the findings of the first large-scale genetic screening of MND genes in New Zealand. A total of 184 MND participants were genetically screened to identify causative variants. Autosomal dominant pathogenic variants were identified in 33/184 (17.9%) participants (24 C9orf72, 9 SOD1), with 12 of these being unaffected members of families with familial MND. One of the participants with C9orf72 repeat expansion was identified as unusual, having tested negative for the expansion by testing of blood DNA during life but whose brain DNA tested positive for pathogenic expansion. Repeat-primed PCR for the expansion revealed an atypical out-of-phase stutter, likely caused by a 2-5 bp insertion or deletion within the expanded allele. Subsequent neuropathological examination of the participant’s brain confirmed a genetic diagnosis of C9orf72. Cumulatively this case highlighted the importance of genetic variability and how this can affect common diagnostic techniques.
The genetic screen also uncovered fourteen variants of uncertain significance within the cohort in which the DCTN1:c.279+1G>C, p.? variant was further investigated. RNA sequencing of dermal fibroblasts from a participant with this variant confirmed the emergence of a novel -108 bp cryptic splice site within the mature mRNA of the proband. This altered splicing coupled with gene conservation and protein domain analysis is suggestive of re-classification of the variant to pathogenic, however, more investigation into whether DCTN1 haploinsufficiency is sufficient to cause MND.
Cumulatively, the results in this thesis are the first to help to explain some of the genetic heterogeneity present within a New Zealand MND cohort and could provide valuable insight into future directions and considerations for the genetic testing of MND in New Zealand
A Virtual Patient Platform for Drug Delivery to Children with Chronic Lung Disease
No full textAsthma is one of the most common chronic respiratory conditions in children, with substantial variation in disease burden across populations. In New Zealand, pediatric asthma hospitalizations remain disproportionately high, highlighting the need for improved diagnosis and treatment strategies tailored to the unique anatomical and physiological characteristics of developing lungs. Aerosolized therapies are a cornerstone of asthma management, yet their effectiveness depends heavily on airway geometry and airflow distribution—factors that are markedly different in children compared to adults.
This thesis presents a virtual pediatric lung modeling framework to investigate how developmental anatomy and bronchoconstriction patterns influence airflow and aerosol deposition. A 1D particle transport and deposition model was first developed and validated using adult airway geometries, for which experimental deposition data are available. In parallel, a statistical shape model (SSM) was constructed from adult end-expiratory CT scans to reflect lung morphology under conditions comparable to free-breathing pediatric MRI. These adult-derived models provided a foundation for generating anatomically realistic pediatric lungs by scaling lobar geometry and airway structure according to child-specific MRI dimensions.
A cohort of child-specific airway models was generated by perturbing shape modes from a statistical adult lung template, capturing key morphological differences such as reduced lung volume, globular shape, and narrower conducting airways. These models were used in conjunction with a one-dimensional structure-function model to simulate airflow and particle transport across a wide range of particle sizes (0.01–10 µm) under baseline, homogeneous, and clustered constriction conditions.
Findings demonstrate that pediatric airway geometry significantly increases bronchial deposition for 1–5 µm particles due to enhanced impaction. Homogeneous constriction consistently shifted deposition proximally and increased resistance, limiting deep-lung delivery. In contrast, clustered constriction localizing to a subset of distal branches, produced spatially heterogeneous deposition, reducing delivery to the obstructed lobe while enhancing sedimentation in adjacent, unobstructed regions.
These outcomes align with clinical imaging studies reporting persistent, lobar ventilation defects in asthma and underscore the limitations of uniform airway scaling in pediatric modeling. This work emphasizes the importance of anatomically accurate, regionally resolved simulations to better understand therapeutic aerosol behavior in children. The developed framework offers new insights for optimizing inhaled drug delivery and supports the advancement of personalized treatment strategies in pediatric asthma care
CPR manikin diversity for BLS education: Current status mapped by an international cross-sectional survey and steps to reach health equity
No full textBackgroundCertain community groups receive less bystander Basic Life Support (BLS). To improve that it was proposed to include in BLS training manikins representing diverse groups, as in current BLS training most manikins are white, lean and male/flat-chested. However, instructors attitudes about the use of diverse manikins and their distribution worldwide are unclear.MethodsA cross-sectional survey was distributed in international resuscitation networks and national resuscitation councils. Data from participating organisations and manikin characteristics used for BLS training were analysed, and differences between countries from different income classification were assessed.ResultsAfter de-duplication and removal of incomplete responses, data of 133 organisations from 43 countries from six continents reporting on 5,364 manikins were analyzed. Most organisations (55%) use only white, male/flat-chested, lean manikins. Non-white manikins were the most commonly used diversification (33% of participating organisations). Only 20% of organisations use female manikins. Greater diversification is thought to enhance realism in training, promote inclusivity, and allows participants to be more aware of real-world situations involving diverse patient populations. Barriers described were high costs, low awareness towards the need of manikin diversity, institutional resistance to changes, and limited evidence on the impact of diversification.ConclusionThe vast majority of reported adult and pediatric CPR manikins are white, male/flat-chested, and lean, and thus lack diversification. Almost one-fifth of respondents indicated to put a bra on a "standard" manikin to simulate a female manikin. Research into diversified manikin use, how to overcome barriers, and its impact on educational and clinical outcomes are needed
On Regular Polytopes of Rank 3
No full textIt is proved that if a finite group G is generated by three involutions α, β and γ, such that α and γ commute, and the orders of the products αβ and βγ are greater than 2, then the generating set {α, β, γ} makes G the automorphism group of a regular 3-polytope if and only if the intersection ⟨αβ⟩ ∩ ⟨βγ⟩ contains no non-trivial normal subgroup of G, and the intersection ⟨α, β⟩ ∩ ⟨β, γ⟩ is not an elementary abelian subgroup of order 4. This criterion complements a theorem by M. Conder and D. Oliveros (J. Combin. Theory Ser. A, 2013, v. 120, no. 6, pp. 1291–130