Crafting and refining high-dose-rate brachytherapy treatment plans for cervical cancer is a time-consuming process. In recent years, BRIGHT was developed, an AI-based automated treatment planning method that provides not just one, but a set of optimized, patient-specific treatment plans, each with a different trade-off between objectives of interest. BRIGHT's plans are optimized using protocols that define guidelines on the delivered doses. In this thesis, we explore an alternative approach using dose-response models. These models provide insights into the estimated outcomes and risks of a treatment plan. Additionally, they offer great adaptability by including external patient characteristics. Currently, the use of these models remains limited to a feedback role. However, we instead include these models in BRIGHT to optimize them directly. Using one Tumor Control Probability (TCP) model and five Normal-Tissue Complication Probability (NTCP) models, we designed and tested several dose-response objective formulations and optimization techniques. We found that with the current models, these new objectives are insufficient as a replacement for BRIGHT's protocol-based coverage and sparing objectives. The produced plans have greatly improved dose-response outcomes but fall short in protocol compliance. Extending the existing objectives rather than replacing them proved more favorable. Average model improvements around 0.004 for NTCP are observed among the best coverage-sparing plans satisfying the protocol. Additionally, by sacrificing some sparing in the protocol-satisfaction range, improvements around 0.005 are possible for TCP and NTCP. Moreover, these dose-response-focused plans show distinct differences in their dose distribution favoring the dose-response targets compared to regular BRIGHT. Ultimately, the improvements we obtained are only marginal, and the clinical implications of this are unclear. The covariates of the models used in this thesis mostly overlapped with BRIGHT's objectives and did not fully extract their potential. Nevertheless, this thesis proves that the concept is viable and builds a foundation for this technique for when more and better dose-response models become available.

Unsupervised machine learning is a complex and abstract topic, posing challenges for student comprehension. Considering the considerable growth of relevance the topic of machine learning has seen in the past years, teaching it effectively has become ever-so important. Analogy-based teaching approaches offer a potential solution by mapping unfamiliar machine learning concepts to familiar real-world ideas. This paper investigates how analogies can improve the understanding of unsupervised learning, a rather relevant field within machine learning. Contributions include a collection of analogies for teaching unsupervised ML, an expert-based evaluation of these analogies’ quality, and a student-centred assessment of analogy-based teaching.
The findings from the expert evaluation show a consensus on the effectiveness of several analogies and highlight which analogies might be less effective. The findings from the student assessment suggest that the analogical explanations are more effective than 'generic' explanations and suggest that students have a higher satisfaction while learning through analogies.
We conclude that well-crafted analogies can enhance student understanding in unsupervised machine learning. The study’s insights can guide educators in integrating analogies to make unsupervised learning more accessible.

This study examines the effect of analogies on conceptual understanding of machine learning (ML) loss functions, and the motivation to learn in first-year bachelor computer science students. For a set of 10 ML loss functions, analogies were generated and evaluated by 15 experts. 3 of these analogies were subsequently tested with 22 students. The results show no conclusive evidence for improvement in understanding and motivation to learn. The study outlines a general strategy for evaluation of analogies on student understanding and motivation. The study further provides 10 expert-rated analogies, 3 of which have been tested with students.

This research paper looks into the influence of domain specificity on the understanding and motivation of first-year computer science students learning different concepts in supervised machine learning. Two types of domains were chosen for the analogies, the general domain and the gaming domain, the latter being the more specific one. These were evaluated in two phases. First, experts rated the analogies based on different metrics. Then, a user study was carried out using A/B testing to measure knowledge gain and motivation when exposed to the analogies. Results from the user evaluation show no statistically significant differences in terms of understanding for domain-specific or general analogies. Motivation, similarly show little difference when comparing both domains. The findings suggest that if analogies are helpful when it comes to understanding a topic, as long as the learner knows the domain, they do not play a big role.

Machine Learning is becoming a standard part of Computer Science curriculums at universities. This paper aims to contribute to the education of Machine Learning in Computer Science, specifically through teaching concepts related to Gradient Descent (GD) through analogies. First, concepts related to Gradient Descent were collected through the use of academic textbooks, and analogies were created based on the definitions found. These analogies were then evaluated by experts, scoring the analogies on Target Concept Coverage, Mapping Strength, and Metaphoricity. The analogies that scored highest on a mean average were then used in an A/B survey distributed amongst Computer Science students that had not followed any Machine Learning course. One group was given the concept definitions, the other both the definitions and the analogies. The learning proficiency was measured, and no statistically significant result was found. In the end, this research explores the possibilities of creating analogies to explain machine learning concepts, and provides a modular framework for evaluating quality and measuring effectiveness of analogies.

Machine Learning education faces significant challenges due to the abstract and mathematically-complex nature of fundamental models, such as Multilayer Perceptrons (MLPs). This paper investigates the effectiveness of conceptual metaphors and analogies as pedagogical tools to improve novice learner's understanding of key MLP concepts. Using large language models, we generated a set of analogies for core MLP topics. These analogies were then evaluated by experts to assess their quality, followed by a user study with novice learners employing a between-subject A/B test comparing analogy-based explanations to formal definitions. Although the study found no statistically significant improvement in knowledge gain or engagement that could be attributed to analogy-based explanations, trends suggest potential benefits in learner confidence and motivation. The research contributes a curated set of expert-evaluated analogies for ML education and discusses methodology limitations and directions for future work. This study highlights both the promise and complexity of integrating analogy-based teaching approaches into ML education.

Consumer-grade fitness trackers can produce unreliable physiological data due to sensor errors. The same holds for cycling data from Wahoo Fitness, where heart rate (HR) and power readings are essential for training and performance analysis. This thesis presents a prediction-based anomaly detection framework tailored to multivariate time-series cycling data. The approach reframes anomaly detection as a personalized physiological HR prediction problem. We define anomalies as deviations between measured sensor values and their predicted values, based on contextual activity metrics (e.g., power, cadence, speed, altitude, and gradient) and user-specific embeddings. The system combines ordinary differential equations (ODEs) modeling heart rate dynamics with machine learning techniques to capture non-linear, non-stationary, and individualized relationships. The model not only detects implausible values but reconstructs them with physiologically consistent alternatives. Compared to reconstruction-based methods, which are mostly used for anomaly detection in time series data, this physiologically grounded approach better differentiates between normal variation and true anomalies. Experimental results demonstrate effective identification and correction of HR and power anomalies, contributing to improved data quality and reliability in wearable fitness applications.

A deeper understanding of Multiple Sclerosis (MS) symptom progression is required for diagnostic accuracy and patient care. Remote monitoring through smartphones can provide continuous insights in the well-being of MS patients. This research aims to explore differences between MS and healthy typing behavior through collected typing behavior data (keystroke dynamics) on smartphones. The main contribution of this work was finding patterns within typing data that distinguish healthy typing behavior from MS with unlabeled time series data, except the diagnosis. A dataset of three combined studies was provided, which included 182 patients suffering from MS and 335 healthy subjects. An autoencoder was fitted on healthy samples to learn the temporal patterns in healthy typing behavior. Then, an anomaly detection method leveraged both the reconstruction loss and embeddings of these samples to determine whether MS samples deviated from healthy behavior. MS subjects were evaluated against healthy subjects through the ROC-AUC on both sample level and subject level. An initial hold out dataset of 20\% of the subjects was left out for validation. The overall performance was poor (AUC 0.6-0.8), because we aimed to find one general method for all MS subjects. Key limitations were the large variance in healthy typing behavior, relatively short window lengths, and difficulty in quantification of the training step of the autoencoder. Despite the difficulty of evaluating the performance of our suggested method, the described methods provide interesting monitoring possibilities for neurologists to enhance MS diagnosis. These monitoring possibilities could extend towards monitoring the effect of treatment, which is the first step towards personalised healthcare in MS.

The intensive care unit (ICU) is a hospital department where critically ill patients requiring organ support or intensive monitoring are admitted. Nowadays, the care provided in an intensive care unit has advanced so that more patients are being discharged alive. Advances in ICU care have increased patient survival rates, yet ICU admission is still associated with high morbidity and mortality both during and after the stay.

Identifying patients at high risk of complications during ICU admission is crucial. Even though there has been an increasing focus on predictive models, making accurate predictions with the data currently available remains challenging. A non-conventional parameter that appears to have promising value is heart rate variability (HRV), which reflects the fluctuations in time intervals between consecutive heartbeats and can be derived from the electrocardiogram. We investigated whether heart rate variability was able to predict ICU mortality and ICU length of stay. We employed a machine learning approach, assessing three models for each outcome. We used nested cross validation to estimate the performance and optimize and select the final models. Data from 468 adult patients admitted to the ICU for 48 hours or longer were analyzed and nine HRV measures were calculated. Two HRV measures, the power in the high frequency band and the standard deviation of 5-minute average RR intervals, showed a significant difference between the ICU survivors (n=398) and ICU non-survivors (n=71). While individual HRV measures had limited predictive power for ICU mortality, combining HRV with clinical features improved performance. The best performing model was an eXtreme Boosting Gradient classifier that used clinical features in combination with three HRV measures (power in the high frequency band, power in the low frequency band and the ratio between the power in the low frequency and high frequency band) achieving an AUC of 0.76. The models predicting ICU length of stay performed poorly, with the best model achieving a mean absolute error of 5.07 days. These findings suggest that HRV, when combined with clinical features, has potential in predicting ICU mortality, though further research is necessary before clinical implementation.

Monitoring ICU survivors after discharge is equally important, as half of these survivors suffer from Post Intensive Care Syndrome, which negatively impacts their quality of life and increases their healthcare needs. A team of researchers from the ICU conducted a pilot study establishing the feasibility of monitoring fifteen ICU survivors using eHealth, but a larger clinical trial was needed to assess feasibility on a larger scale. Therefore, we aimed to prepare for the ICU Recover Box 2.0 study. We evaluated the results and challenges of the first pilot. The study protocol was revised to include more participants and the use of new smart technology, specifically the Corsano CardioWatch, replacing the non-CE marked devices from the first pilot, which required 24/7 researcher availability. Preparations included finalizing the application to the Medical Ethical Review Committee and thinking out the study logistics, resulting in a comprehensive plan for the conduct of the study. Key takeaways from this process include the recognition that research is an iterative process of
continuous learning, that the purpose of the study must be carefully considered with ethical concerns in mind, and that the increasing importance of data requires careful planning for its security, processing and storage. The METC application has been submitted. Once approved, the study can proceed on a well-prepared basis.

In reinforcement learning, the ability to generalize to unseen situations is pivotal to an agent’s success. In this thesis, two novel methods that aim to enhance the generalizability of an agent will be introduced. Both of the methods rely on the idea that the diversity of a replay buffer increases an agent’s ability to generalize. The first utilizes the agent’s exploration strategies to reach interesting states. The second aims to reach further using an additional goal-conditioned agent. Both methods demonstrate improved adaptability without relying on domain-specific knowledge and show promising results.

In human-AI agent interactions, providing clear visual or textual explanations for the agent's actions and decisions is crucial for ensuring successful collaboration. This research investigates whether having the visual explanations displayed only on-demand, instead of having them consistently shown as the baseline, has an impact on the human supervisor's level of confidence and satisfaction with the AI agent. Therefore, a case study of 40 participants was conducted to explore this hypothesis and the participants were divided into 2 groups, one interacting with the on-demand condition, and the other with the baseline one. Through questionnaires, the participants' capacity and moral trust in the robot, the explainable artificial intelligence satisfaction, and the disagreement rate with the robot's decisions have been collected. Demographic data was gathered from the participants to explore whether their background could impact the collaboration. This data included the participants' gender, age, education, gaming experience, risk propensity, trust propensity, and utilitarianism. The resulting statistical analyses indicated no significant differences between the baseline and the on-demand conditions concerning trust and explanation satisfaction. This suggests that the overall collaboration was not primarily impacted by the frequency of visual explanations requested on demand. Although the results implied a high satisfaction with the interaction, further studies with more diverse user groups are recommended. Overall, this research reinforces the importance of transparency in decision-making processes during the collaboration between an AI agent and a human supervisor.

The integration of robots in human-robot teams, particularly in high-stakes environments like firefighting, requires effective communication and decision-making to ensure safety and efficiency. This study explores the impact of adding contrastive explanations to feature attributions in robot explanations on human-robot teamwork during firefighting simulations. Contrastive explanations aim to improve human understanding by highlighting why a robot chose one decision over another using allocations of variables. The experiment involved 40 participants, divided into two groups, each interacting with either the baseline or contrastive version of the robot in the simulated environment. Results indicate that contrastive explanations significantly increased participants' capacity trust in the robot, though they did not significantly affect moral trust. Additionally, the results showed a lower satisfaction level with the explanations given by the robot. The disagreement rate between human decisions and robot actions was lower in the contrastive group, suggesting possible enhanced understanding and agreement with the robot's decisions. These findings underscore the potential of contrastive explanations to enhance trust and collaboration in human-robot teams, paving the way for more effective integration of robots in critical operations. Future research should focus on larger sample sizes and explore the inclusion of contrastive decisions made by the robot alongside explanations to further validate these findings.

The Algorithm Selection Problem (ASP) presents a significant challenge in numerous industries, requiring optimal solutions for complex computational problems. Traditional approaches to solving ASP often rely on complex, black-box models like random forests, which are effective but lack transparency, and they often fail to balance performance with interpretability. This paper investigates the performance-interpretability trade-off for the ASP, specifically focused on Optimal Decision Trees (ODTs) as recent innovations have made the use of ODTs more viable. We compare ODTs against 4 other tree-based models, using 11 different datasets. We show there is no apparent tradeoff between performance and interpretability for ODTs which have been trained using an instance cost-sensitive approach, as they achieve comparable performance to a Random Forest Regressor while maintaining interpretability through multiple orders of magnitude fewer leaf nodes.

Survival analysis predicts survival functions that give the probability of survival until a given time. Many applications of survival analysis involve health care, which requires interpretability of the models used to predict the survival function. Provably optimal decision trees have shown to be an interpretable alternative to so-called black box models. However, these algorithms often choose estimators that are fast, yet not necessarily most accurate. Moreover, the objective functions of optimal decision tree algorithms tend to make (possibly incorrect) assumptions about the survival function. In this paper, we tackle both problems. We implement the iterative Breslow estimator in an already existing optimal survival tree algorithm in order to iteratively improve the Nelson-Aalen estimator. This approach has great potential, as we show by using it on artificial datasets, but we do not see an improvement in accuracy on real world data. To eliminate the assumptions made by the objective function, we implement the Integrated Brier Score objective, which causes a significant improvement on training accuracy. However, we see no improvement on out-of-sample accuracy

Decision trees are valued for their ability to logically and transparently classify data. While heuristic methods to compute such trees are efficient, they often compromise on accuracy, prompting interest in Optimal Decision Trees (ODTs), which have the best misclassification score for a given tree size limit and training dataset. The dynamic programming (DP) approach for ODTs has shown improvements over alternatives such as mixed-integer or constraint programming. That being said, it requires further improvements to handle exponential runtime scaling with the depth of the tree and the number of features of the dataset. Leveraging modern hardware, such as multiple CPU cores, offers a promising solution to improve efficiency. This paper proposes a multithreading method for DP ODTs which we apply to STreeD specifically. We introduce a shared memory model and determine which components of the original program can be made local to the threads. We investigate whether it is more efficient to start multithreading at the root of the search tree than near its leaf nodes. We find the former to be superior, resulting in faster runtimes and less shared resource access. Empirical evaluations against the state of the art demonstrate better runtimes, particularly beneficial for large datasets. Finally, thread scaling analyses reveal substantial speed-ups, exceeding 2.5 times with four threads, highlighting our approach's effectiveness for computationally-intensive tasks.

Survival analysis is a branch of statistics concerned with studying and estimating the expected time duration until some event, such as biological death, occurs. Survival distributions are fitted based on historical data, where some instances are censored, meaning that the actual time of the event is not known precisely. Survival trees extend on the classical statistical methods developed and can capture complex non-linear relations between the variables by recursively splitting the instances by generated rules and fitting a different survival distribution in each leaf. Moreover, decision trees are desirable models due to their interpretable nature. We extend existing optimal survival tree methods by considering Cox Proportional Hazard models in each leaf node, which allows us to find more complex yet interpretable relationships than existing methods. The experiments show that our model outperforms state-of-the-art methods for creating survival trees, SurTree, OST, and CTree, especially in determining the relative risks between out-of-sample observations while generating significantly smaller trees.

In the pursuit of employing interpretable and performant Machine Learning models, Decision Trees has become a staple in many industries while being able to produce near-optimal results. With computational power becoming more accessible, there has been increasing progress in constructing Optimal Decision Trees. It guarantees optimal solutions with respect to different metrics within a given size limit on training data while requiring a smaller number of nodes and becoming more viable to compute on real-world data. However, non-linear metrics, which are very effective when evaluating trees on imbalanced datasets, still represent a challenge regarding runtime performance and scalability. Previous approaches generate the Pareto Front of the set of possible solutions, an expensive operation in computing the optimal tree. To address this gap, we introduce a novel merging algorithm of two Pareto Fronts using convex hulls, offering better pruning and leading to an increase in scalability. The experiments show a significant improvement in runtime of almost 10\% on bigger datasets and higher-depth trees using the F1-score metric, with the potential to be applied to other convex metrics.

Counterfactual explanations (CEs) are emerging as a crucial tool in Explainable AI (XAI) for understanding model decisions. This research investigates the impact of various factors on the quality of CEs generated for classification tasks. We explore how inter-class distance, data imbalance, balancing techniques, the presence of biased classifiers, and decision thresholds influence CE quality. To answer these research questions, we conduct experiments on various datasets, classification models and counterfactual generators. The datasets include the MNIST and GMSC dataset. The models include well-established models like MLP and Random Forest, along with the novel NeuroTree model. The generators include the method proposed by Wachter et al. and the REVISE method. We evaluate how different factors affect CE quality by performing an extensive experimental analysis. Our findings demonstrate that increasing inter-class distance degrades CE quality, particularly explanation plausibility. Data imbalance showed minimal impact, while balancing techniques yielded a slight improvement in CE plausibility, especially for the minority class. Classifiers biased towards specific subgroups resulted in lower CE quality for those subgroups. We observed limited evidence for a consistent amplification effect of decision thresholds. This research utilizes various datasets and classification models, including the novel NeuroTree model. Our findings contribute to XAI by providing insights into factors affecting CE quality and highlighting areas for future development, particularly regarding fairness and handling imbalanced data.

Optical Character Recognition (OCR) is a pivotal technology used to extract text information from images, finding wide-ranging applications in document digitization and medical records management. The integration of machine learning has ushered in an era of swift and precise OCR models. Broadly, OCR comprises two key components: detecting the bounding boxes around text instances and recognizing the characters within them. Presently, prevailing OCR models are primarily intricate two-stage systems necessitating real-time operation on remote servers. Nevertheless, end-to-end models exhibit superior performance from a data utilization perspective. There exist scenarios where offline models prove indispensable, such as in environments with restricted internet access or locales with stringent data privacy and security requirements.

This project delves into various end-to-end models, leveraging the PaddleOCR end-to-end model as a foundational reference to devise a compact OCR model tailored for edge devices. Through meticulous optimization of the backbone architecture and the introduction of diverse Feature Pyramid Network (FPN) structures within the stem network, we achieved a remarkable reduction in model size, down to 19MB. This represents a substantial advancement, constituting merely one-tenth of the original PaddleOCR end-to-end model's footprint.

By leveraging an extensive database and conducting a series of fine-tuning experiments specifically tailored for end-to-end OCR tasks involving curved text images, the model exhibits an impressive precision rate of 47.3% and an f-score of 45.3%. This achievement highlights the effectiveness of the customized loss function relative to the original model, despite its reduced size. Notably, this performance is comparable to certain end-to-end models with larger backbones. Furthermore, an Android demo has been carefully developed to demonstrate the model's capabilities on mobile devices, achieving an average processing time of 433 milliseconds per image.

Intrusion detection systems (IDSs) are essential for protecting computer systems and networks from malicious attacks. However, IDSs face challenges in dealing with dynamic and imbalanced data, as well as limited label availability. In this thesis, we propose a novel elastic gradient boosting decision tree algorithm, namely Elastic CatBoost Uncertainty (eCBU), that adapts to concept drifts and copes with label scarcity by using a novel sequential uncertainty estimation method. We compare our method with state-of-the-art techniques on synthetic and real-world datasets and show that it achieves comparable accuracy and higher robustness to limited label availability in intrusion detection tasks.