Objective
The primary aim of this study was to develop and validate a machine learning prediction model for respiratory deterioration in mechanically ventilated Intensive Care Unit (ICU) patients. The secondary aim was to identify physiological parameters associated with respiratory failure during mechanical ventilation.

Methods
Two distinct prediction models were developed using data from ICU patients admitted to the Leiden University Medical Centre (LUMC) between 2018 and 2023. Patients receiving invasive mechanical ventilation (IMV) for at least 48 hours with a PaO2/FiO2 ratio below 40 kPa were included and allocated to COVID training, COVID test, or non-COVID test sets. Model 1 predicts respiratory deterioration within six hours after switching from controlled to assisted ventilation. Model 2 is an hourly updating model predicting respiratory deterioration occurring more than six hours after this switch. XGBoost models were cross-validated on the COVID training set to identify the optimal observation windows and prediction horizons, after which feature selection and hyperparameter optimisation were performed. Model 1 was optimised for the area under the receiver operating characteristic (AUROC) and Model 2 for the area under the precision-recall curve (AUPRC). Discriminative performance, generalisability, and clinical utility were evaluated on the COVID and non-COVID test sets.

Results
A total of 296 patients were included in the COVID training set, 78 in the COVID test set, and 755 to the non-COVID test set. For Model 1, a one-hour observation window was selected. The most important features were the mean fraction of inspired oxygen (FiO2), propofol infusion rate, and peripheral oxygen saturation (SpO2). This model achieved an AUROC of 0.78 on the COVID test and 0.76 on the non-COVID test set. For model 2, a two-hour observation window and a six-hour prediction horizon were selected, with the SpO2/FiO2 ratio as the most important input feature. This model achieved an AUPRC of 0.05 on the COVID test set and 0.03 on the non-COVID test set.

Conclusion
Model 1 demonstrated moderate discriminative performance but limited clinical utility at relevant operating points. Model 2 showed very limited predictive value, primarily due to extreme class imbalance. Consequently, neither model is currently suitable for clinical implementation. With larger datasets and more advanced modelling techniques, Model 1 may have the potential to become a clinically useful decision support tool to support decisions on switching from controlled to assisted ventilation.

Estimating the effect of an intervention on an outcome is a central challenge across science and society. In medicine, we may ask whether a drug effectively treats a disease, and in economics, whether a new policy reduces unemployment. Estimating such effects from data, a process known as causal inference, is essential but inherently difficult because it often relies on untestable assumptions to ensure unbiased identification of treatment effects. A key example of such an untestable assumption is the absence of unmeasured confounding, meaning that no hidden variable influences both the treatment and the outcome. When this assumption fails, something which we cannot directly verify, treatment effect estimates may become biased. This ultimately can lead to untrustworthy conclusions and, in the worst case, unsafe decisions, such as prescribing the wrong drug to a patient. The central question of this dissertation is therefore whether we can develop methods for safer causal inference that either detect violations of its underlying assumptions or remain robust when those assumptions are violated.
In Part One, we address the first aspect of detecting violations of causal identification assumptions. We focus on settings with data from multiple sources, such as hospitals or locations, where distributional shifts naturally occur. Under specific independence conditions on the causal mechanisms driving these shifts, we first present a nonparametric test to falsify the assumption of no unmeasured confounding. To obtain these results, we introduce a novel technique utilizing hierarchical causal graphical models. Thereafter, we focus on improving the statistical efficiency of this test, which is achieved by reformulating the independence condition using parameterized linear models. Finally, we extend the hierarchical modeling approach to other identification settings, specifically by testing the validity of mediators and instrumental variables used in two additional common identification strategies.
In Parts Two and Three, we develop methods that instead are robust when causal identification assumptions are violated.We revisit two commonly occurring problem settings when doing causal inference and demonstrate that it is possible to develop methods that either remove the need for, or rely on, weaker and more plausible assumptions than those traditionally made. In the first setting, we study the problemof augmenting randomized trials using external data to improve efficiency in treatment effect estimation. Typically, such approaches rely on a transportability assumption that relate the populations underlying the trial and external data. But when this transportability assumption is violated, integrating external data can introduce substantial bias. To address this, we propose a novel and efficient estimator that incorporates external data and show that this estimator improves inference on the average treatment effect while guaranteeing that it never performs worse, and sometimes performs better, than the estimator that relies solely on trial data.We further adapt this estimator to learn heterogeneous treatment effects within the trial population and show that similar safety guarantees hold for this problem.
In the second setting, we examine the evaluation of treatment allocation strategies using Qini curves. Standard methods for estimating Qini curves assume no interference between treated units, meaning that the treatment of one unit does not affect others. However, when interference is present, these Qini curves can be misleading and lead to incorrect evaluation of treatment allocation strategies.We therefore propose multiple estimators to handle the interference, specifically in settings where units within a cluster may affect one another but not units in other clusters.We identify a bias-variance trade-off in these estimators and, through both theoretical and empirical results, provide practical guidance on how practitioners can choose among them. The dissertation concludes with a discussion of broader considerations, limitations of the presented research, and potential directions for future work.We find that it is indeed possible to make causal inference safer by detecting assumption violations and reducing reliance on untestable assumptions. Nonetheless, many open and important questions remain, offering promising avenues for further research on this topic.

Osteoarthritis (OA) is a prevalent musculoskeletal disease, and radiographic assessment remains the standard for diagnosis and grading. However, expert grading is subjective and intensity-based automated methods are sensitive to imaging variability. As a potential solution to these problems, landmark-based approaches are worth exploring. Landmark-based representations of bone geometry offer an alternative to pixel-based inputs, reducing sensitivity to imaging artifacts and emphasizing structural variation. This thesis compares four landmark encodings (raw x,y coordinates, Procrustes-aligned points, pairwise distances, and polar coordinates) and evaluates them using both linear dimensionality reduction (PCA) and nonlinear generative modeling (VAEs) on hip radiographs from a publicly available dataset. We evaluate reconstruction fidelity, latent space traversal, correlation with clinical outcomes, and classification performance. Results show that raw point coordinates provide a strong baseline, often matching or outperforming more complex encodings in classification, while alternative representations improved interpretability but not discriminative power. PCA preserved clinically meaningful variability, whereas VAEs underperformed in this unsupervised setting. These findings suggest that landmark annotations already contain sufficient information for supervised OA tasks, while more advanced models may be needed for unsupervised or generative applications.

Combining data from Randomized Controlled Trials (RCTs) is a widely used method to estimate causal treatment effects. In order to combine data, the property of transportability, under which different covariate vectors exhibit similar treatment benefit, must hold between the RCTs. However, differences in study design, execution, and the underlying effect modifier distributions can violate transportability which could in turn lead to estimating incorrect causal treatment effect estimates. This thesis addresses the challenge of validating transportability between multiple RCTs and identifying subsets of RCTs between which transportability holds. Our contributions include studying a linear regression-based framework for testing transportability between multiple RCTs and a clustering-based approach for identifying transportable RCT subgroups. Through simulations and analysis of real-world RCTs concerning corticosteroid treatment for Community-acquired pneumonia (CAP), we evaluate the power, robustness, and limitations of our proposed framework.

Osteoarthritis is a chronic joint disease in which the protective cartilage between bones deteriorates over time, leading to pain, stiffness, and reduced mobility. Diagnosis is a time-consuming and somewhat subjective process. To address this challenge, machine learning techniques can be applied. However, training supervised models on medical images is often challenging because of the limited availability of labeled training data. Self-supervised methods, which pretrain models to learn useful features without labels, offer a potential solution to this issue. In this paper, we explore the use of Generative Adversarial Networks (GANs) as a pre-training step for osteoarthritis diagnosis. The first step is the training of a GAN on a semi-public dataset of x-ray images. In the second stage, we explore different strategies for fine-tuning the discriminator model to diagnose osteoarthritis. Our experiments suggest that while GAN-based pre-training offers slight improvements over purely supervised approaches, the performance gains remain modest.

Self-supervised learning (SSL) is a promising approach for medical imaging tasks by reducing the need for labeled data, but most existing SSL methods treat each scan as an isolated sample and overlook the fact that patients often have multiple radiographs taken over time. These longitudinal sequences—multiple scans of the same hip acquired at different visits—encode the natural progression of osteoarthritis (OA) and thus could enrich representation learning. In this study, we evaluate whether incorporating temporal information from these longitudinal radiographic sequences into SSL pretraining yields more transferable representations and leads to improved downstream classification of hip OA severity. We focus on a temporal contrastive task (Contrastive Predictive Coding, CPC), which learns to predict future scan representations from earlier ones, and compare it to a SimCLR-based pretraining that treats each radiograph independently. We also investigate a multitask framework that combines both objectives — either by sequentially pretraining with CPC then SimCLR, or by interleaving the two tasks. Experiments on the Osteoarthritis Initiative (OAI) dataset for binary classification of KL-grade severity show that CPC alone does not surpass SimCLR-based pretraining. However, both the sequential and interleaved multitask approaches significantly improve classification accuracy over either single-task method. These findings demonstrate that even though temporal prediction by itself isn’t sufficient — combining temporal and within-scan contrastive learning can yield stronger models for hip OA severity assessment.

Supervised learning approaches have proven to be useful in diagnosing Osteoarthritis from X-ray images, aiding professionals in an otherwise time-consuming and subjective process. However, in the medical field, labeled data is scarce. For this reason, we investigate a contrastive self-supervised approach, SimCLR, capable of learning useful representations from unlabeled data. Specifically, we explore a core component of this method – the data augmentation techniques. While these augmentations are highly effective in introducing variability in conventional image datasets, they are too aggressive for medical images, often altering their semantic meaning. In this paper, we implement custom anatomy-aware augmentation techniques, which aim to preserve the main region of interest needed for a diagnosis. We evaluate these anatomy-aware augmentations including Gaussian blur, Contrast enhancement, Random resized crop, and Random erasing, against their classical counterparts by training multiple encoders based on different combinations of those augmentations. The findings of our study have shown that utilizing this anatomy-aware approach for all data augmentations a model uses does not lead to a significant improvement in its performance. However, selective use of anatomy-awareness on geometric-based approaches seems to show promising initial results.

Causal inference, particularly the estimation of the Conditional Average Treatment Effects (CATE), is necessary for understanding the impact of interventions beyond simple predictions. This study analyzes the influence of key hyperparameter choices, specifically maximum tree depth and minimum leaf size, on the accuracy and generalization of CATE estimates derived from honest and adaptive causal trees. The research explores how these hyperparameters affect the bias-variance trade-off and the model's tendency to overfit or underfit across various simulated data scenarios and a real-world dataset.

The results reveal that optimal hyperparameter configurations are dependent on the data characteristics, such as dimensionality, noise levels, and the complexity of the true causal effects. Honest causal trees demonstrate a better performance in high-dimensional and noisy environments due to their effective variance control. Conversely, in simpler, low-noise settings or complex CATE structures, adaptive causal trees or baseline models frequently achieve better results by reducing bias. The study also highlights the challenges of using moderately sized datasets, where the sample splitting limitations can lead to higher estimation errors. This work provides thorough suggestions for hyperparameter selection, emphasizing the fact that tuning based on the underlying characteristics of the data is needed for achieving the best CATE estimates possible.

Estimating the Conditional Average Treatment Effect (CATE) with neural networks adapted for causal inference, like TARNet, is a promising approach, yet the impact of model architecture on performance remains underexplored.
This paper systematically investigates how the depth and width of TARNet affect the CATE estimation in diverse simulated data environments. The research investigates two central questions: how TARNet's performance varies across data regimes (e.g., confounding strength, sample size), and how its optimal architecture changes in response to these conditions.
A comprehensive set of simulation-based experiments is conducted using the CATENets framework, isolating and varying factors such as sample size, feature dimensionality, confounding strength, and the presence of noise. The results demonstrate that deeper architectures generally yield better performance in complex or high-dimensional scenarios, whereas narrower networks are preferable in small-sample or high-noise settings due to their regularizing effect. Furthermore, the findings suggest that there is no universally optimal architecture. The best configuration depends on the specific characteristics of the data. The study concludes with practical recommendations for architecture selection based on the experiments conducted.

This study tackles an important issue in evaluating the reliability of confidence intervals in causal forests by examining how data characteristics and hyperparameters influence actual coverage rates compared to theoretical benchmarks. Using synthetic data sets with polynomial treatment effects, Sobol sampling, High-Dimensional Model Representation (HDMR), and comprehensive grid searches, the study assesses causal forest performance in different data contexts.

A primary discovery is the identification of a practical limit for reliable confidence interval coverage: When the sum of confounders and effect modifiers exceeds 4, coverage rates drop considerably below 80%, even for simple treatment effect functions. This limitation remains steady despite substantial increases in computational resources.

The examination of hyperparameters revealed that the most influential parameters are the maximum tree depth and the balance tolerance in splits, which demonstrate substantial changes in performance, both of which performed best at their maximums (unlimited and 0.5, respectively). Other key suggestions involve increasing the training data fraction per tree from 0.45 to 0.5, keeping the minimum impurity decrease threshold at 0.0, and utilizing at least ≈ 2400 trees to meet theoretical expectations.

In addition, this paper did not identify any noteworthy interaction between tree count and sample size. As a result, both of these characteristics can be optimized independently of each other.

These findings provide systematic guidelines for practitioners to assess when causal forest confidence intervals are reliable and how to optimize them, bridging the gap between theoretical guarantees and practical performance.

Causal Multi-task Gaussian Processes (CMGPs) provide a Bayesian approach for estimating in-
dividualized treatment effects by modeling potential outcomes as correlated functions. However,
they struggle under high-dimensionality and treatment imbalance, leading to overfitting and unre-
liable uncertainty estimates. This study examines two failure modes: poor generalization in high-
dimensional spaces and overconfident predictions in low-overlap regions. To address these, two
data-aware enhancements are proposed: an overlap-adaptive kernel that scales similarity based on
local treatment density, and a regularized prior that down-weights unstable features using marginal
treatment effect variance. Evaluations on synthetic data and the IHDP benchmark show improved
effect estimation, credible interval calibration, and robustness in challenging settings. These find-
ings highlight practical strategies for enhancing CMGPs in real-world causal inference tasks.

Interventional Normalizing Flows (INFs) are a recently proposed method for estimating interventional outcome distributions from observational data. A central component of this approach is the nuisance flow, whose function is to estimate the propensity score and the conditional outcome distribution. INFs are claimed to be doubly robust, meaning they can yield valid estimates even if only one of these components is correctly specified. This study investigates the practical limits of this robustness by asking two questions: (1) How do interventional estimates behave when nuisance flow components are entirely misspecified? and (2) How sensitive are these estimates to more realistic imperfections such as suboptimal hyperparameters or injected noise? Through experiments on four benchmark datasets with varying levels of confounding and distributional complexity, we find that INFs remain robust under low-confounding conditions even when both nuisance components are broken. However, in highconfounding settings, even partial misspecification can cause estimates to degrade substantially, undermining the doubly robust property. These results highlight the importance of carefully validating nuisance components and suggest that the theoretical guarantees of INFs may not always hold in practice.

Self Supervised Learning (SSL) has been shown to effectively utilise unlabelled data for pre-training models used in down-stream medical tasks. This property of SSL enables it to use much larger datasets when compared to supervised models, which require manually labelled data. Medical classification tasks often require the identification of patterns inside a small Region Of Interest (ROI) known to be relevant for radiographic diagnosis. This contrasts standard image classification tasks, which generally rely on broader patterns. To guide a model in learning such anatomically relevant features, we investigated the hip osteoarthritis classification performance of a ROI-guided Masked Autoencoder (MAE) with a Convolutional Neural Network (CNN)-based architecture. Unlike conventional MAEs, which learn latent features by reconstructing randomly masked images, our alternative uses generated anatomical landmarks to exclusively mask the ROI or background. Contradicting similar research on Vision Transformer (ViT)-based MAEs, random masking outperformed our ROI-guided alternatives, revealing a fundamental difference in what drives performance for the two architectures, and guiding future research on more sophisticated ROI-guided masking strategies. The code is available on GitHub: https://github.com/Jasperdetweede/AnatAMAE/

Statistical distribution alignment methods for domain adaptation assume similar class distributions across domains, but this assumption cannot always be guaranteed in medical imaging data. This research investigates the effect of cross-domain class imbalance on statistical distribution alignment in unsupervised domain adaptation for medical image classification. Our experiments demonstrate that statistical distribution alignment using MMD performs reliably under mild domain shifts but struggles when both severe cross-domain class imbalance and complex domain shifts are present. To address this, we implement class-conditioned domain alignment with a new weighted minibatch sampling method. Under conditions of extreme domain shift and severe cross-domain class imbalance, combining statistical distribution alignment with more complex sampling strategies results in small improvements compared to alignment with random sampling, suggesting that class-conditioned distribution alignment offers limited practical benefits. The model appears robust to label noise, but since the performance gains are tiny, the choice of sampling strategy could have limited influence on overall performance. In our experiments, we employ the CHECK and OAI hip X-ray datasets to investigate binary osteoarthritis classification under varying levels of domain shift and cross-domain class imbalance.

Causal inference methods are often used for estimating the effects of an action on an outcome using observational data, which is a key task across various fields, such as medicine or economics. A number of methods make use of representation learning to try to obtain more
informative feature representations, which are then used for effect estimation. However, such feature representations can introduce confounding bias in the results if they lose confounding information contained within the original features. In this work, we evaluate an existing metric for measuring confounding bias in representations and use this metric to study two representation learning methods in terms of their biases in settings with low overlap between treated and control populations. We show that the metric is a suitable measure for the amount of confounding bias in a representation and representation learning methods that minimise this bias lead to better average treatment effect estimation in our experiments.

The field of causal inference provides a variety of estimators that can be used to find the effect of a treatment on an outcome based on observational data. However, many of these estimators require the unconfoundedness assumption, stating that all relevant confounders are observed within the data. This assumption is quite strict and many real-life problems would violate it, due to confounders being too expensive, immoral, or abstract to observe. To weaken this assumption, causal sensitivity analysis attempts to quantify the level of hidden confounding and use it to create a possible bound on the causal effect. This study compares a well-established Marginal Sensitivity Model (MSM) with a newly proposed f-sensitivity model. Given f-sensitivity is relatively new, the current approaches for computation can be inefficient or non-deterministic. A new method of computing the f-sensitivity bound is proposed, which can lead to closed-form solutions in some estimator-specific cases. The differences between the MSM and f-sensitivity models are outlined and illustrated with examples, from which a set of guiding questions is created for researchers to decide between the two sensitivity models. All of the code required to reproduce this study is located in this GitHub repository.

This research investigates the use of Causal Multi-task Gaussian Process (CMGP) for estimating the individualized treatment effect (ITE) of low versus high Positive End-Expiratory Pressure (PEEP) regimes on ICU patients requiring mechanical ventilation. The study addresses the complexities of determining ITE due to the inability to observe counterfactual outcomes and the confounding bias in observational studies. By employing Conditional Average Treatment Effect (CATE) estimators, such as S-Learner, T-Learner, and CMGP, the research evaluates the impact of different PEEP settings on patient survival across varied patient characteristics. The precision of these estimators is assessed using simulated data, real-world observational data from the MIMIC-IV dataset, and an external RCT dataset. The findings of this study are inconclusive, highlighting the need for further research to refine these methods and explore larger, more balanced datasets.

The severity of hip osteoarthritis is measured a.o. by the minimal distance between the femoral head and the acetabular roof in an X-ray image. However, the whole joint space profile might be a more accurate estimator, since it would include irregularities in the bone surface. These irregular bulges (osteophytes) on the bone surface are one of the signals that a person might have OA. Thus the stage of OA might be better estimated automatically by having this data in the joint space profile instead of just using the minimal joint space.

For this joint space profile, the distance between the femoral head and the acetabular roof needs to be calculated. Therefore, the positions of these parts in the hip joint are required to be know. These can be retrieved from e.g. a segmentation mask.

One way of calculating the distance in a joint is to use a radial projection. A radial projection is a way of projecting points from a curved space to a plane by projecting lines from a central point along increasing angles.

In this paper, we investigate how the joint space profile can be segmented most accurately from a radial projection originating from the center of the femoral head by several comparing noise filtering and edge-finding algorithms. After which is shown that a custom algorithm based on the theory behind edge detection in noisy images works most reliably and accurately.

There are still multiple points of improvement for this algorithm. The femoral head can be segmented more accurately than the acetabular roof, the segmentation of the latter could be optimized by detecting the brightest line (peaks) instead of the most sudden change (steepest gradient) in the X-ray image as the edge for the femoral head. The algorithm could be further improved by taking care of local outliers off those edges.

In conclusion, this paper compares multiple ways of segmenting the joint space of the hip joint. The best-performing algorithm could in the future be used in an assisting tool for doctors to highlight important irregularities and measurements in the hip joint space.

Mechanical ventilation with positive end-expiratory pressure (PEEP) is a critical intervention for patients in intensive care units (ICUs) with acute respiratory failure. Identifying the optimal PEEP level is challenging due to conflicting evidence from studies comparing low and high PEEP regimes. This research explores machine learning methods for estimating individualized treatment effects (ITE) in ICU patients on different PEEP levels using the observational MIMIC-IV dataset. Various conditional average treatment effect (CATE) estimators, including S-, T-, and DR-learners, are applied to control for confounders and identify PEEP effects on patient subgroups. This research aims to compare the performance of the aforementioned CATE estimators, with a focus on the doubly-robust (DR) learner, and determine which one is best suited for causal inference in this context. The DR-learner offers increased resilience to model errors since it integrates two models. Simulations using mean squared error (MSE) show the DR-learner performs well with confounded data and differing linear response functions between control and treatment groups. However, when looking at the performance on the MIMIC-IV dataset, the predictions are unstable, failing to reliably identify the optimal PEEP for increasing patient survival. This trend is also observed in a randomized controlled trial (RCT) dataset, with the area under the Qini curve (AUQC) close to zero, indicating difficulties in identifying the effects of PEEP settings. Despite promising simulation results, real-world application shows limitations in these machine learning methods for optimal PEEP identification.

Osteoarthritis is a degenerative disease that affects the aging population by degrading the cartilage in the joints. The early and accurate diagnosis of this disease is key to effective treatment. For an early and accurate diagnosis of this disease, clinicians often use X-ray imaging. This allows medical professionals to manually measure the joint space width (JSW) in X-rays images to determine the progression of the disease. This method however proves to be both time-consuming and variable based on the professional. This research addresses the automation of the measurement of the JSW for the hip, using deep learning techniques, to improve precision and efficiency. The automated measurement of the JSW is challenged by variations in the imaging conditions across different clinical settings. To address these discrepancies and keep a good performance, domain adaptation techniques are used to counter these domain shifts to ensure a consistent JSW segmentation across different imaging domains. The study investigates whether a specific domain adaptation technique can enhance the accuracy and robustness of deep learning models specifically for femur segmentation in X-ray images across different datasets. A base deep learning model is developed for femur segmentation, and supervised domain adaptation is applied. The study compares the performance of the adapted model with the base model across two different datasets. Results indicate that supervised domain adaptation does not significantly improve the model’s robustness and accuracy in femur segmentation among two different datasets. These unexpected findings suggest that incorporating domain adaptation techniques may not always lead to a more reliable and efficient diagnosis of osteoarthritis, reducing the manual workload for clinicians.