Nitrous oxide ( N2O ) is a widely used inhalational analgesic and anesthetic across clinical settings including obstetrics, pediatrics, emergency medicine, and dentistry. However, its global warming potential (273 times that of CO2 over a 100-year horizon) and documented occupational health risks make its unmitigated clinical use increasingly difficult to justify within contemporary sustainability frameworks.

This thesis provides a comprehensive, multi-dimensional evaluation of N2O management in healthcare, integrating a systematic technology review, an institutional case study at Leiden University Medical Center (LUMC), and a clinical deep-dive into obstetric practice.

Chapter 1 maps the global landscape of N2O delivery, scavenging, and destruction technologies across a dataset of 60 devices from 19 manufacturers. Findings reveal that 78.3% of catalogued systems provide zero active greenhouse gas mitigation, with true catalytic destruction present in only 11.7% of devices. Significant geopolitical stratification limits technology access in lower-income healthcare systems, while regional regulatory divergence drives fragmented clinical adoption.

Chapter 2 quantifies seven years of institutional N2O consumption at LUMC (2019–2025), modeling cumulative unmitigated emissions of 103,637 kg CO2-eq across four departments. The sharpest emissions growth was driven by the 2023 introduction of N2O -based cryoanalgesia for pediatric pectus excavatum repair. Substantial discrepancies between theoretical and real-world carbon footprints were identified, attributable to point-of-use catalytic destruction systems and infrastructural transitions. Critical regulatory compliance gaps in ambient occupational exposure monitoring were documented.

Chapter 3 analyzes 403 obstetric labor episodes at LUMC. Despite modest objective pain reduction (mean VAS decrease of 1.63 ± 1.48 points), patient and provider satisfaction scores remained high (7.31 and 7.76 out of 10, respectively), illustrating the well-documented analgesic paradox of N2O . The overall analgesic conversion rate to epidural or remifentanil analgesia was 18.35%, rising to 25.37% among primiparous patients, supporting risk-stratified clinical routing recommendations.

Collectively, these findings demonstrate that sustainable N2O governance cannot be achieved through isolated end-of-pipe solutions. Instead, institutions require a coordinated strategy encompassing decentralized cylinder-based supply, workflow-matched abatement technologies, risk-stratified clinical protocols, and governance structures that integrate frontline clinical expertise. The Dutch regulatory framework, combining occupational health law, sustainability covenants, and professional clinical standards, is identified as an instructive model for international healthcare systems seeking to reduce direct anesthetic gas emissions without compromising patient care.

Numerical information shapes how people interpret real-world events, evaluate claims, and make decisions. As Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) systems are increasingly used for open-domain information seeking, their ability to reason reliably over quantities has become critical. Yet, even when models retrieve relevant evidence and produce fluent chain-of-thought reasoning, they still often reach incorrect conclusions by failing to reason about numbers in the right contexts. This work investigates numerical scoping errors, an underexplored reasoning failure where a model incorrectly binds a quantity to the wrong referent, time-frame, unit, population, condition, or source. In high impact settings, a single mis-scoped number can change the meaning of a conclusion, turning a statistic into a misleading financial signal, an unsafe medical recommendation, or a false public claim. This makes numerical scoping errors an important reasoning failure to investigate. When performing numerical fact-checking (FC) and question-answering (QA), this work identifies scoping errors as a common and consequential failure mode in generated LLM reasoning traces across tasks. In human analysis, scoping errors appeared in 44.7% of numerical FC and 26.8% of numerical QA traces, with 35.3% and 81.8% of these errors, respectively, judged to contribute to incorrect final answers. To detect these failures at scale, this work develops a hybrid LLM-as-a-Judge pipeline that decomposes reasoning traces into quantities, verifies whether each quantity is correctly scoped based on its source type, and aggregates these quantity-level judgements into trace-level labels. The pipeline achieves moderate agreement with human annotations, but can be expensive to leverage. To explore more efficient alternatives, the hybrid pipeline's scoping detection signal is distilled into smaller, conservative student models which process traces around 212 times faster. Finally, this work evaluates whether scoping-aware reward signals can improve downstream performance in a parallel test-time scaling setting, where the system chooses among multiple generated reasoning traces, producing modest improvements, especially when combined with correctness-based selection. Overall, this work shows that numerical reasoning failures are not merely errors of calculation but errors of context. By making these contextual failures visible and measurable, this work takes a step toward numerical reasoning systems that can support high-impact, real-world decisions not only because they calculate correctly, but because they understand what their numbers mean.

Vascular phantoms are artificial vessel models that provide controlled and repeatable environments for medical research, imaging validation, device testing, surgical planning, and training. To improve their physiological relevance, these phantoms should not only reproduce vascular geometry, but also show mechanical behaviour that is representative of blood vessels. In particular, small-scale compliant vascular phantoms should be able to deform under pressure-driven flow. However, fabricating small open channels from soft materials while maintaining geometrical accuracy, tunable stiffness, and measurable compliance remains challenging.

This thesis investigates the fabrication and experimental characterization of small-scale flexible vascular phantoms created using masked stereolithography 3D printing. Straight cylindrical channels were printed using Elastomer-X resin with different lumen diameters and wall thicknesses to determine a printable geometry range. The geometrical accuracy of the printed channels was evaluated using microscopy, while the internal lumen geometry was further inspected using ink-filled channels. Tensile tests were performed to determine the Young's modulus of the printed material under different exposure and post-curing conditions. By varying the printer exposure time and UV post-curing duration, different stiffness configurations were obtained.

The pressure-flow behaviour of the printed channels was investigated using water as the working fluid. Baseline measurements were used to correct for pressure losses in the setup, after which the channel-only hydraulic resistance was calculated. The effective diameter was then estimated from the hydraulic resistance using the Hagen--Poiseuille relation. The results showed that a lumen diameter of 0.7~mm with a wall thickness of 0.25~mm could be printed as the smallest consistently open channel suitable for flow experiments. Tensile testing showed that the Young's modulus could be tuned between approximately 1.05 and 1.25~MPa by changing the fabrication conditions. Pressure-flow measurements showed a decrease in hydraulic resistance with increasing mean pressure for several channels, corresponding to an increase in effective diameter. The softer channels generally showed a larger relative effective diameter increase than the slightly stiffer channels.

These results demonstrate that masked stereolithography can be used to fabricate small-scale flexible vascular phantoms with tunable stiffness and measurable pressure-dependent behaviour. Although the effective diameter was determined indirectly and uncertainties remain due to local lumen narrowing, measurement noise, and baseline correction, the study provides an important experimental validation step towards compliant 3D printed vascular phantoms for pressure-flow applications.

The Internet connects organizations across finance, government, education, and other sectors, forming a global network of interdependent systems. Although this connectivity enables organizations to provide services, coordinate operations, and communicate with external users, it also increases their exposure to cyber attackers. Attackers exploit connected networks for financial gain, political objectives, or operational disruption, while often hiding their identity through distributed infrastructure such as botnets. These botnets are obtained either by renting existing infrastructure or by manually compromising large numbers of computers. To maximize the value of this infrastructure, attackers reuse the same botnets across multiple targets within a single campaign. Consequently, organizations in the same industry sector, which often expose similar network protocols and services, face related malicious activity from the same attacker-controlled infrastructure.

To defend themselves against such attackers, organizations deploy intrusion detection systems (IDSs). An IDS continuously monitors the incoming and outgoing network traffic, typically represented as packets, and inspects this traffic for signs of malicious behavior. IDSs commonly rely on two primary detection methods. Signature-based detection checks whether the packets match known attack signatures. Anomaly-based detection identifies traffic that deviates from expected or normal network behavior. Together, these methods help organizations detect known attacks, unusual behavior, and suspicious traffic patterns within their own networks.

Although individual IDS deployments provide network defenses, attackers exploit the limitations of local detection. For example, signature-based detection fails when attackers use evasion techniques such as packet fragmentation, where malicious traffic is split into multiple smaller packets that appear harmless when inspected separately. More broadly, a single IDS observes only the traffic directed at its own network. As a result, it sees suspicious activity as isolated local events and lacks the wider context needed to determine whether the same attacker infrastructure is targeting other organizations. Therefore, an IDS alone is not sufficient to identify broader attack campaigns that span multiple networks.

The inability of individual organizations to observe campaign-level activity creates a need for greater cross-organizational visibility. When the same attacker-controlled infrastructure targets organizations in the same sector, shared patterns appear across their network logs. Correlating these observations helps organizations determine whether suspicious activity is isolated or part of a coordinated campaign. However, obtaining such visibility requires multiple organizations to compare information derived from their logs. Directly sharing raw network logs is impractical because logs contain sensitive information, including credentials, internal IP addresses, authentication data, proprietary operational details, and other organization-sensitive information.

The hesitation in sharing raw network logs creates the need for a computational framework that enables secure correlation between organizations. Private Set Intersection (PSI) provides a cryptographic basis for this capability by allowing parties to compute common elements across private sets without revealing non-matching elements. PSI has been extensively studied and has been extended to Multi-party Private Set Intersection (MPSI), which supports n participants. A relevant variant for collaborative threat intelligence is Threshold MPSI, where an element is reported if it appears in at least t out of n private sets. In this thesis, the proposed TMPSI-based solution determines whether a suspicious IP address appears in a threshold number of organizations' network logs and alerts the participating organizations to possible coordinated attack activity. Organizations that have not yet observed the same activity can use the threshold result to proactively strengthen their defenses. In this way, the proposed framework transforms isolated IDS deployments into a collaborative threat intelligence system that preserves privacy, revealing attack campaigns that remain hidden from individual organizations while preserving the confidentiality of local network logs.

Building energy performance is largely determined during the early conceptual design stages. However, energy evaluation is often postponed until after key architectural decisions—such as massing and façade configuration—have been finalized, limiting opportunities for meaningful optimization. This study addresses this gap by demonstrating a Building Information Modeling (BIM)-based comparative approach for integrating energy analysis into early design stages. Focusing on residential villas in Abu Dhabi, UAE, three residential forms were selected for testing: compact, L-shaped, and U-shaped designs. Each form included multiple iterations representing the design progression process, namely the base building form, façade refinement (glazing and shading adjustments), and spatial layout and/or orientation adjustments. The analysis incorporated variations in glazing ratio and building orientation to evaluate their combined impact on Energy Use Intensity (EUI). Among the tested options, the L-shaped configuration achieved the lowest EUI (79 kWh/m2), representing the best-performing option, followed by the compact (81 kWh/m2) and U-shaped (82 kWh/m2) configurations. The results also confirm that orientation remains an important factor even after façade refinement. Overall, the findings suggest that BIM-based energy analysis is most effective when applied early as a comparative and iterative design-support tool.