70040 research outputs found
Sort by
Legal Compliance and the Open Texture of Law
No full textThe law is often vague and ambiguous, especially when applied to novel and unusual cases. Legal scholars have referred to this as “the open texture of law.” Legal compliance is seldom straightforward, requiring interpretation before conceiving and designing mechanisms for compliance. Organizations find themselves having to plan for and ultimately tackle compliance under uncertainty. This policy editorial discusses closure in legal compliance in the context of the open texture of law, using the example of the EU Artificial Intelligence Act. This example is of particular concern for information systems (IS) research and practice. This policy editorial aims to offer some guidance in this complex area
Perinatal and infant outcomes after assisted reproductive technology treatment for endometriosis alone compared with other causes of infertility: a data linkage cohort study
Full text linkObjective: To evaluate whether perinatal and infant outcomes differ between singleton births after assisted reproductive technology (ART) in women with endometriosis alone and those with other causes of infertility. Design: Population-based data linkage cohort study. Subjects: A total of 29,152 ART-conceived singleton births from 24,116 mothers, 2010–2017, New South Wales, Australia. Exposure: Endometriosis, identified from the Australian and New Zealand Assisted Reproduction Database, hospital admissions, and dispensed medication records. The causes of infertility were categorized as follows: endometriosis alone; endometriosis plus other cause(s) of infertility; infertility other than endometriosis; and unstated cause of infertility. The endometriosis alone group was further classified using International Classification of Diseases, Tenth Revision, codes (N80.0–N80.9) into superficial, ovarian, deep, and other endometriosis. Main Outcome Measures: Perinatal and infant outcomes, included preterm birth (<37 weeks), very preterm birth (<32 weeks), small for gestational age (SGA), large for gestational age, admission to neonatal intensive care unit, perinatal death, and infant hospitalization up to 2 years of age. Generalized estimating equations were used to investigate independent associations between endometriosis and study outcomes. Results: Of the 29,152 ART-conceived singleton births, 19.9% (5,806/29,152) were from mothers with a diagnosis of endometriosis. Among these, 23.8% (1,379/5,806) were from mothers with an endometriosis alone diagnosis, and 76.2% (4,427/5,806) were from mothers with endometriosis plus other cause(s) of infertility. Three quarters (21,795/29,152) of births were from mothers without endometriosis and 5.3% (1,551/29,152) were from mothers with an unstated cause of infertility. After adjusting for maternal age at the time of birth, parity, ART treatment characteristics, gestational hypertension and diabetes, smoking, and socioeconomic status, there was no overall association between endometriosis and perinatal and infant outcomes. However, compared with women without endometriosis, those with deep endometriosis had a higher risk of preterm birth (adjusted relative risk, 1.75; 95% confidence interval, 1.12–2.75) and SGA (adjusted relative risk, 1.58; 95% confidence interval, 1.05–2.37). Conclusion: Reassuringly, perinatal and infant outcomes are generally comparable for ART-conceived infants born to mothers with endometriosis alone and those with other causes of infertility when considered as a singular disease entity. Larger studies are needed to confirm the differential risk associated with endometriosis phenotypes; however, for patients with deep endometriosis undergoing ART, the risks of preterm birth and SGA may be increased. Clinicians should be aware of these potential risks
Planning for Scale-Up and Sustainability: A Multi-site Process Evaluation Protocol for a Novel Intervention for Survivors of Childhood Brain Cancer
Full text linkComplex interventions often fail to sustain widespread reach at a population level, despite demonstrating clinical effectiveness during piloting and trial evaluation. ‘Engage’ is a multi-disciplinary and risk-stratified intervention that is delivered remotely to childhood cancer survivors to promote equitable and improved access to survivorship care. Engage is not a standalone intervention in that it requires careful consideration of how it will be integrated into existing survivorship care pathways. Our study aims to conduct a process evaluation of the Engage intervention as applied to brain cancer survivors (‘Engage Brain’) to further contextualise trial outcomes, and understand what factors contribute to a sustainable, scalable, and successfully implementable intervention. A mixed-methods process evaluation will be conducted as part of the Engage Brain type-1 effectiveness-implementation trial. Data collection will occur across four domains of research: (1) planning, (2) implementation, (3) practice setting, and (4) ecological setting. Data sources will include semi-structured clinical stakeholder interviews, primary care practitioner interviews, transcribed implementation meetings and project log, transcribed nurse consultations, study materials, and administrative/process data. Qualitative data will be analysed using both deductive and inductive thematic analysis, guided by implementation science frameworks such as the updated Consolidated Framework for Implementation Research, which encompasses the Theoretical Domains Framework and implementation outcomes. Quantitative data will be analysed and presented using descriptive statistics where appropriate. Conducting a process evaluation underpinned by implementation science and behaviour change theories will enable the development of a national scale-up framework and improved delivery of sustainable models of care for childhood cancer survivors.
Trial Registration: The Australian and New Zealand Clinical Trials Registry (ANZCTR), https://www.anzctr.org.au, ACTRN12621000590864
Exploring experiences of talk therapies among gay and bisexual men seeking to reduce or abstain from using crystal methamphetamine in the context of chemsex
Full text linkIntroduction: Some gay and bisexual men who have sex with men (GBMSM) who use drugs to enhance sex (chemsex/party and play) may experience harms and seek talk therapies. GBMSM who practice chemsex may not access drug services because of anticipated stigma and the perception that these services lack chemsex expertise. Barriers to services are documented, however, little is known about the service experiences of chemsex engaged GBMSM. Methods: Semi-structured interviews were conducted with 24 participants reporting current practice of sexualised use of methamphetamine and/or gamma hydroxybutyrate. Interviews explored experiences of counselling and psychology services, participant's treatment goals and challenges. Data were transcribed verbatim and analysed in NVIVO14 with a qualitative description methodology. Results: Most in our study sought to reduce the frequency of methamphetamine use and used methamphetamine only in sexual contexts. When engaging with counsellors and psychologists in alcohol and other drug or mental health services for the general adult population, most censored the sexual drivers and types of sexual behaviours incumbent in their methamphetamine use. Participants' reliance on drugs for sex was spoken about as a major barrier to reducing methamphetamine. Sexual self-censorship within services inhibited participants' abilities to access meaningful support and achieve treatment goals. Discussion and Conclusions: Counsellor and psychologists working with GBMSM around drug use, must ask about context of drug use and sex. Training and supervision around sexual therapies for those working alongside GBMSM who practice chemsex may be beneficial. Research on treatment approaches to support the sexual wellbeing of people who practice chemsex is required
Using open data to create a meaningful crisis dashboard in Australia.
Full text linkAs populations grow and the effects of a changing climate contribute to increasing numbers of disasters of increasing intensity, there is a need to keep the public abreast of disaster information so that they can be empowered to make decisions that will better protect their family, friends, assets and community. It is in this context that this work examines the disconnect between the information the public expect to receive during a disaster and what is made available by the disaster response agencies. It investigates also the experiences the public have when accessing critical information.
This study employed a hypothetical disaster scenario, and a survey conducted in two sequential parts, before and after a demonstration of a prototype crisis dashboard that utilises open emergency data. The approach looked to uncover survey participants' experience of accessing emergency information before receiving a demonstration of the prototype crisis dashboard. The second part of the survey was designed to determine whether the survey participants felt that the prototype crisis dashboard offers an improved approach to accessing emergency information. The quantitative and qualitative responses to the survey were analysed to derive input from the respondents on the information they believe is useful to them and how they would like to access it.
Improving access to emergency information and empowering the community to make informed decisions during an emergency would enhance outcomes for community members by reducing loss of life and asset damage, minimize the associated financial costs, and shorten the recovery phase. It would also alleviate the pressure on emergency services
Ionically crosslinked biohybrid gelatin-based hydrogels for 3D cell culture
Full text linkThe transition from two-dimensional to three-dimensional cell cultures has transformed the understanding of cell physiology and cell–matrix interactions. Extracellular matrix (ECM) mimics tend to fall into either the natural or synthetic categories. Naturally occurring ECM mimics, such as collagen and gelatin, have superior bioactive properties but typically lack tuneability. Conversely, synthetic ECM mimics are highly defined but even with modifications, can lack the bioactivity of natural proteins. Therefore, to take advantage of the potential of both natural and synthetic ECM mimics, a biohybrid ionically crosslinked gelatin hydrogel was synthesised. This was achieved by utilising free amine groups along the gelatin backbone as the basis for a reversible addition − fragmentation chain-transfer (RAFT) reaction. The resulting polymers had tuneable stiffness and enhanced solubility compared to gelatin. The biohybrid gel also showed good biocompatibility, with MCF-7 cells forming larger spheroids when encapsulated within the biohybrid gel when compared to an unfunctionalized polyethylene-glycol (PEG) gel. Furthermore, due to the ionic crosslinking in the biohybrid gel, spheroids can be retrieved by digesting the matrix using 10 × phosphate-buffered saline (PBS). Retrieved cells were shown to be viable which allows for the potential of downstream analysis. Thus, this study highlights the potential of hybrid gelatin–PEG hydrogels for 3D cell culture
Development of low carbon concrete with high cement replacement ratio by multi-response optimization
No full textThis study develops three new Low Carbon Concrete (LCC) mix designs with characteristic cylinder compressive strengths of 32 MPa (C32), 25 MPa (C25), and 20 MPa (C20). By using a Taguchi design of experiment (T-DoE) model and combined it with Grey relational analysis (GRA) and Principal component analysis (PCA) for multi-response optimization, sixteen trial mixes employing supplementary cementitious materials (SCMs) to replace 80 % to 95 % of ordinary Portland cement (OPC) were tested. Three factors namely, OPC replacement percentage, ground granulated blast-furnace slag (GGBFS) to fly ash (FA) ratio, and silica fume (SF) to binder percentage were considered. Optimization results led to three LCC mix designs with 80 %, 85 %, and 90 % OPC replacement. Their compressive strength, split tensile strength, flexural strength, elastic modulus, and slump were evaluated. Confirmation tests showed that the 80 %, 85 % and 90 % OPC replacement mixes respectively satisfied requirements for C32, C25, and C20 concretes. Carbon footprint study showed that the LCC mixes led to significant reduction of carbon footprint when compared with OPC concrete. Finally, microstructure analysis was conducted to study in the microstructure characteristics of the LCCs
Characterising and optimising the electrical responses of healthy and degenerating retinas to improve artificial vision
No full textEarly retinal prostheses successfully elicited visual percepts in individuals with profound vision loss, however, the percepts were impractical for critical tasks like reading and navigation. Acknowledging that these devices relied on constrained and unoptimised stimulation regimes, this thesis aimed to better understand, characterise, and optimise the electrical responses of the output cells of the retina – the retinal ganglion cells (RGCs) – under both healthy and degenerating conditions.
We first studied stimulation of the neuron-dense human central retina by developing a large-scale computational model of foveal RGC populations. Unlike predecessors, the model exhibited eccentricity-dependent anatomical accuracy and fovea-specific characteristics. We used this model to reproduce response patterns akin to percepts reported by implant recipients and to optimise responses based on electrode configuration and stimulation waveform. Our findings indicated that a multi-return hexapolar configuration could elicit robust single-RGC activation in 56% of locations tested. We observed a negligible difference between stimulation waveforms in terms of activation thresholds and response confinement, proposing that this indifference was a limitation of the model’s lack of a retinal network.
Subsequently, we sought to study the electrical responses of RGCs using patch clamping. To enable this, we first developed a template-subtraction-based pipeline for removing arbitrary electrical artifacts caused by various stimulation waveforms and recovering stimulus-embedded neural activity. The novel technique demonstrated a true positive rate of 89%, false discovery rate of 3%, and d-prime of 3.11. Our ensuing experimental findings demonstrated that phase width had a statistically significant impact on temporal response profiles and charge efficiency. Waveform shape was heavily confounded by other parameters but significant under the right conditions. Although stimulus-response strengths were similar between mid-to-late-stage degenerating retinas and intact and synaptically-blocked healthy retinas, degeneration severely impaired the utility of electrically-elicited network-mediated activity. Whole-cell profiling indicated that RGC excitability was not impaired by retinal degeneration but suggested possible mutation-dependent differences in intrinsic electrical characteristics.
Altogether, these results suggest that, at least for an epiretinal configuration, direct electrical activation of RGCs is a feasible and more reliable therapeutic strategy throughout retinal degeneration than network-mediated activation. Whilst retinal protheses will likely need to cater to the unique context of subject-specific remodelling, optimised stimulation waveforms and configurations may better enable functional artificial vision
Towards Trustworthy Artificial Intelligence through Causal Inference
Full text linkThe primary goal of modern artificial intelligence (AI) is to build trustworthy systems. However, current AI approaches predominantly rely on correlation-based statistical learning, which, while effective in many applications, struggles to capture the causal mechanisms underlying real-world data. This limitation undermines the ability of AI systems to provide reliable reasoning, particularly in dynamic or complex environments. To address this, equipping AI systems with causal inference capabilities is essential, as it enables them to learn causal relationships rather than mere correlations, thereby laying a stronger foundation for truly trustworthy AI.
This thesis aims to enhance the trustworthiness of AI systems by systematically integrating causal inference capabilities. We follow a progressive approach to tackle increasingly complex challenges in causal effect estimation. First, we propose a novel contrastive learning framework to estimate individual treatment effects (ITE) under the assumption of unconfoundedness. We then relax this assumption with a disentangled learning framework using instrumental variable (IV) representations to address unobserved confounding. Under the same assumption, we further explore methods to estimate ITE when valid IVs are absent. Lastly, we extend our study to high-dimensional image data, introducing a causally diffusion-based method to generate counterfactual samples, improving the robustness of prompt learning in vision language models.
The structure of this thesis is organized as follows: The Introduction elaborates on the necessity of integrating causal inference into contemporary AI research, highlights challenges in current approaches, and defines our research objectives. The Literature Review provides an in-depth overview of the foundational principles of causal inference, synthesizes the two major frameworks of causal inference, summarizes its mainstream algorithms, and highlights the latest advancements in leveraging causal inference to empower AI. The Technical Innovations section details our progressive technical approaches to addressing challenges such as confounding bias, hidden confounders, instrumental variables generation and high-dimensional counterfactual generation. The Conclusion summarizes the research findings and proposes potential future research directions. The Appendix furnishes rigorous proofs underpinning the theoretical contributions of this thesis. Through this comprehensive exploration, we aim to advance causal inference in enhancing the reliability and practical applications of AI models in dynamic and uncertain environments
Securing Video Recognition Systems: Mitigating Adversarial Attacks and Establishing Resilient Defenses
Full text linkDeep Neural Networks (DNNs) have significantly advanced video recognition systems (VRSs) in various applications, such as face recognition, anomaly detection, and autonomous driving, largely due to their precise classification capabilities. However, DNN-based video recognition models are prone to adversarial video attacks, which pose a critical security risk to VRSs. Such attacks subtly alter the input videos in ways that are typically undetectable to humans, causing these models to produce incorrect results. For example, subtly altered stop signs can mislead DNNs in autonomous vehicles, risking catastrophic outcomes.
To mitigate adversarial attacks, we propose two defense frameworks, SecVID and ViDToken—each defends against adversarial attacks from a unique angle.
SecVID, the first contribution of this thesis, as a correction-based pre-processing defense, uses discretization-enhanced video compressive sensing in a black-box pre-processing module, transforming videos into a sparse domain to disperse and neutralize perturbations. While SecVID’s discretized compression disrupts perturbation continuity, its reconstruction process minimizes adversarial elements, causing only minor distortions to original videos. Although not completely restoring adversarial videos, SecVID significantly enhances their quality, allowing accurate classification by SecVID-enhanced video classifiers and preventing adversarial attacks.
On the other hand, we propose ViDToken, an innovative pre-inference adversarial video detection framework, as the second contribution of this thesis. ViDToken utilizes latent tokens to detect adversarial video without the need for model access, modifications, or known adversarial examples. Additionally, we introduce Representative Token Selection (RTS), a new strategy that selects the most indicative token of the video, enhancing ViDToken’s efficiency. Finally, we propose a frame-replication that significantly boosts ViDToken against adaptive sparse attacks.
We conduct comprehensive evaluations of SecVID and ViDToken against state-of-the-art adversarial video attacks, and the results consistently demonstrate the strong defense capabilities of both frameworks. Designed for high-risk scenarios, both SecVID and ViDToken address trade-offs such as additional training for their defense modules and the moderate increase in VRS inference times