Applied Machine Learning in Healthcare: Discrimination Risks, Prediction Stability, and Lifecycle Evaluation

By: Ioannis Bilionis
Supervisors: Dr. Carlos Castillo & Dr. Luis Fernández-Luque

Date: October 30, 2026 - 15:00 h
Room: 55.309

Abstract

The rapid adoption of Artificial Intelligence (AI) in healthcare has amplified concerns around data quality, algorithmic fairness, and the reliability of Machine Learning (ML) systems across their lifecycle. Despite significant advances in fairness-aware ML, existing evaluation approaches remain largely centered on predictive performance and static fairness metrics, providing limited insight into how training data composition, prediction instability, and model updates affect equitable outcomes across patient populations. This thesis addresses these gaps through a data-centric and lifecycle-aware set of frameworks and empirical analyses for developing equitable and trustworthy clinical AI.

First, it identifies critical gaps in publicly available health datasets, highlighting the underrepresentation of behavioral, mental health, and sociodemographic variables essential for fairness-aware modeling. Building on this foundation, the thesis demonstrates that dataset balance alone does not guarantee fairness, leveraging Systematic Arbitrariness (SA), a measure of prediction instability, as a novel lens in healthcare to reveal hidden disparities across patient subgroups. It then proposes a Discrimination Risk Assessment Framework that provides an actionable assessment of training data composition, enabling the detection of subgroup-specific risks prior to model deployment through the joint analysis of performance, stability, and training value. Finally, the work extends fairness-aware model evaluation to the post-deployment phase, showing that model updates under real-world data shifts can degrade stability and exacerbate inequities even when accuracy is preserved, and introduces a multidimensional monitoring framework to mitigate these risks.

Overall, these contributions establish a comprehensive approach to clinical ML that integrates data quality, fairness, and robustness. Although the proposed methodologies are empirically validated using representative chronic disease and pediatric diabetes case studies, the underlying frameworks are intended to provide a general foundation for discrimination-risk assessment and lifecycle-aware monitoring across a broader range of healthcare AI applications. 

Efficient Reinforcement and Imitation Learning in Infinite-Horizon Markov Decision Processes

By: Antoine Jean Jacques Nathan Moulin
Supervisor: Dr. Gergely Neu

Date: December 30, 2026 - 16:00 h
Room: TBC

Abstract

This thesis develops statistically and computationally efficient methods for large discounted infinite-horizon Markov decision processes (MDPs) under structural assumptions. For online reinforcement learning, it develops a variant of regularized optimistic value iteration that does not rely on any contractive property, making it especially compelling under function approximation. This method applies to standard structural assumptions like tabular and linear mixture MDPs. Another algorithm further adds optimistic augmentation, a data-dependent soft truncation, for efficient learning in linear MDPs. Both provide regret and sample complexity guarantees. The latter method also handles adaptively changing rewards and imitation from expert features when further environment interaction is available. For fully offline imitation learning, the thesis introduces a primal-dual method that assumes structure in the value functions of policies generated by the learner rather than in the expert policy, providing finite-sample guarantees without assuming that the expert belongs to a known policy class. Together, these results show how structure makes reinforcement and imitation learning tractable.

Advancing the Understanding of Congenital Heart Diseases through Computational Modeling of the Fetal Cardiovascular System

By: Maria Inmaculada Villanueva Baxarias
Supervisors: Dr. Bart Bijnens, Dr. Gabriel Bernardino Pérez & Dr. Patricia García Cañadilla

Date: November 3, 2026 - 12:30 h
Room: 55.309

Abstract

Congenital heart diseases are one of the major causes of infant mortality, affecting approximately 1% of live births. During fetal life, shunts redistribute blood flow and help preserve organ perfusion, partially compensating disease-related alterations and potentially masking them, complicating prenatal diagnosis and severity assessment. The evaluation of fetal health and disease diagnosis relies mainly on imaging techniques, which can only provide a partial characterization of fetal hemodynamics, as they cannot directly quantify pressure or oxygen saturation, and are further limited by spatial resolution, fetal motion, and angle dependence. In this thesis, we advanced the understanding of congenital heart diseases through computational models of the fetal cardiovascular system on different physiological scales. We developed a closed-loop 0D model of the fetal circulation, coupled with an oxygen transport model, and a 3D computational fluid dynamic model of the fetal right atrium. First, we showed that in aortic coarctation, ventricular disproportion and ductus arteriosus dilation compensate for aortic narrowing, preserving pressures, organ perfusion, and most velocities; aortic isthmus velocity pattern might aid diagnosis. Second, we demonstrated that the Eustachian valve promotes preferential streaming of oxygenated ductus venosus blood toward the foramen ovale, influencing cerebral oxygen delivery. Finally, we found that fetal transposition of the great arteries remains relatively hemodinamically stable despite marked variability in oxygen distribution, particularly cerebral oxygenation; foramen ovale and ductus arteriosus velocity patterns may indicate worse postnatal outcomes. 

Efficient Computation of Distances Between Markov Chains: Applications to Reinforcement and Imitation Learning

By: Sergio Calo Oliveira
Supervisors: Dr. Anders Jonsson & Dr. Javier Segovia-Aguas

Date: November 6, 2026 - 14:00 h
Room: 55.309

Abstract

Markov chains are the standard model for sequential stochastic systems, and they sit at the heart of the Markov decision process framework used to model sequential decision-making problems. As learning systems are deployed across increasingly large, complex, and often subtly different environments, representation and generalization have become central concerns. Being able to compare stochastic processes is key to reasoning about transfer, robustness, and generalization. Under that premise, it is crucial to understand how similarity between stochastic processes can be quantified in a way that is both theoretically grounded and computationally tractable. The goal of this thesis is to build tools to answer this question by developing efficient methods to compute a theoretically grounded distance between Markov chains. We start by showing that two notions of distance developed in entirely separate communities (optimal transport distances and bisimulation metrics) are equivalent objects under a given cost function. That equivalence lets us reformulate the distance as a linear program over discounted occupancy couplings, add entropy regularization, and derive a Sinkhorn-like algorithm. The resulting method comes with sample-complexity guarantees and improved computational complexity over prior approaches, though it assumes full knowledge of both chains’ dynamics. We then drop that assumption by characterizing the distance through the chains’ marginal occupancy measures, which can be estimated from samples. We propose what is, to the best of our knowledge, the first method for estimating the distance from sample streams alone: a stochastic primal-dual algorithm with accompanying convergence guarantees. Finally, we derive the gradient of the distance with respect to the parameters of the compared chains and propose a general algorithm for fitting a parametric model to a stochastic process. We apply this method to representation learning, state-space compression, and imitation learning from observations. 

Linguistic Structure in Neural Language Models and their Evaluation: From Multilingual Morpho-syntactic Evaluation to Morphology-aware Tokenization for Spanish

By: Alba Táboas García
Supervisors: Dr. Leo Wanner & Dr. Piotr Michal Przybyla

Date: November 12, 2026 - 11:00 h
Room: TBC

Abstract

Despite the success of large language models, common evaluation practices still provide only a partial view of their linguistic competence, especially beyond English. At the same time, tokenization, a foundational component of model design, is usually optimized for statistical efficiency rather than linguistic adequacy, even though it conditions what models can infer from the input. This dissertation explores two complementary ways of bringing linguistic insight back into the loop: linguistically grounded evaluation as a principled way to probe what models learn, and linguistically informed tokenization as a way to improve how they learn. It does so through two empirical studies: first, developing multilingual targeted syntactic evaluation resources for the analysis of morpho-syntactic competence across languages and models; then, building a morphology-aware tokenization strategy for Spanish that integrates explicit word-internal structure into subword learning and assessing its impact on model quality

A World of Possibilities: Representations of Modal Concepts in Infants

By: Linda Elisiana Exaud Mushi
Supervisor: Dr. Luca Bonatti

Date: November 26, 2026 - 10:00 h
Room: 40.035

Abstract

This dissertation examines the developmental trajectory of modal concepts, in particular measuring whether infants have those associated with understanding multiple alternative possibilities. We develop a paradigm geared at investigating whether participants reorient their gaze towards alternative outcomes when an expected outcome fails to occur. Two experiments on adults study the expected response of beings who clearly possess an understanding of alternative possibilities. Two further experiments on infants, with ages ranging from 10 to 36months, apply this method to the population of interest and begin studying the basic parameters to optimize the procedure. We study how oculomotor responses can be indices of processing of multiple alternatives. We find that infants in certain condition exhibit oculomotor patterns qualitatively similar to those of adults, suggesting that they may have an understanding of alternative possibilities, redirecting their gazes to alternatives when their expectations are unfulfilled. 

Investigations Into Early Forms of Logical Reasoning Accross Categories in Infants and Toddlers

By: Daniel Díaz Fernández
Supervisor: Dr. Luca Bonatti & Dr. Ana María Martín-Salguero

Date: November 26, 2026 - 15:30 h
Room: 40.035

Abstract

The origin of the human ability to formulate and test hypotheses remains unknown. Some research suggests that elementary forms of logical reasoning are available in preverbal infants (Cesana-Arlotti et al. 2018, 2022; Ekramnia et al. 2021). Alternative interpretations argue that such results need not assume logical reasoning or representations of multiple possibilities (Leahy & Carey, 2020; Leahy et al., 2022), but may instead reflect simpler perceptual systems for tracking objects (Jasbi et al., 2019). Here, we present an original eye-tracker experimental design to test whether infants' representation of ambiguous scenes, where object identity may be determined by disjunctive reasoning, can be explained by such less committing accounts. We found that the oculomotor signatures of older 1-year-olds and 2-year-olds, but not younger 1-year-olds, resemble those of adults when they disambiguate the location of a target animal by performing a disjunctive syllogism across categories. 

Exploiting Coherence in Rendering: Clustering for Interactive Global Illumination and Gaussian Processes for Material Modeling and Integration

By: Arnau Colom Pasqual de Riquelme
Supervisor: Dr. Ricardo Marqués

Date: November 27, 2026 - 10:00 h
Room: 55.309

Abstract

The synthesis of physically based realistic images is one of the most computationally expensive processes in Computer Graphics due to the costly operations needed to simulate the paths of light. Furthermore, accurately modeling real-world materials that might be used to characterize the scene's objects is challenging because of their highly irregular behavior and their four-dimensional representation.

In the first part of this dissertation, we propose a new rendering framework based on the many-light problem to allow for VPL-based interactive rendering. By means of clustering the points displayed on the screen and the VPLs scattered in the scene, we are able to significantly reduce the number of ray-primitive intersections used to evaluate visibility. We introduce the idea of fuzzy clustering and present two different fully GPU-contained clustering algorithms. The first clustering method, based on K-means, shows how by applying fuzzy clustering, we are able to render artifact-free images without the need for denoisers. However, in most practical applications, K-means becomes too rigid due to its inherent fixed cluster parameterization. To address this limitation, we introduce the QT-clustering method to allow the user to practically select a quality criterion for the clustering algorithm to then refine the proper number of clusters.

The second part of the thesis incorporates Gaussian processes into rendering. First we propose a BRDF modeling method based on Gaussian processes. Contrary to previous works, which require large training times and are dependent on large amounts of data, our approach is shown to train very fast with very few observations while improving the BRDF fitting of state-of-the-art methods. Additionally, we show how the model can be easily adapted to anisotropic materials.

Regarding the offine estimation of the rendering integral, previous works presented a Bayesian Monte Carlo rendering framework based on Gaussian processes. However, these methods are constrained by the prohibitive cost of matrix inversion. This is a problem in rendering, since it is usually of interest to iteratively add samples that provide the maximum amount of information about the incident radiance function, which can only be learned through expensive sampling. In this last work, we present an efficient updating method for the BMC estimator by leveraging Cholesky factorization and propose the construction of a new heuristic for actively sampling the regions of interest. We demonstrate that updating the BMC quadrature by leveraging Cholesky factorization allows for faster updates of the integral estimate, while also showing the potential of a Spherical Harmonics-based heuristics utilizing Gaussian processes for sample selection.

Investigation of Treatments for Depression with Whole-brain Dynamics and Modelling

By: Paulina Clara Dagnino
Supervisor: Dr. Gustavo Deco

Date: December 1, 2026 - 15:00 h
Room: 55.309

Abstract

Depression is a leading contributor to the global burden of disease. Different evidence-based treatments have demonstrated clinical efficacy, yet a deeper understanding of their neural signatures is needed to optimise effects. This thesis investigates functional magnetic resonance imaging (fMRI) brain changes following escitalopram, psilocybin, and mindfulness-based cognitive therapy (MBCT) in depression, as well as meditation in health. Using whole-brain dynamics and modelling as a unifying framework, perturbational, thermodynamic-inspired and dimensionality reduction approaches are applied to characterise the brain’s hierarchical and spatiotemporal organisation. The findings reveal that each intervention is associated with specific brain changes, with convergence across methods supporting the robustness of the results and differences adding new layers of knowledge. Importantly, brain markers are linked to clinical and behavioural outcomes and interpreted within psychological frameworks. Overall, this work contributes to our understanding of how escitalopram, psilocybin, and MBCT modulate neurocognitive functioning in depression and opens avenues for future research.

Improving the Ecological Validity of Synthetic Datasets for Deep Neural Network-based Audio Processing

By: Enric Gusó Muñoz
Supervisors: Dr. Xavier Serra & Dr. Umut Sayin Saraç

Date: December 3, 2026 - 16:00 h
Room: 55.309

Abstract

Deep Neural Network (DNN) based audio processing models often fail to generalize from their test sets to real-world conditions, where performance can degrade substantially. This thesis pursues to reduce this gap by improving the ecological validity of the (often synthetic) data. We develop this idea across three audio processing tasks.

(i) For binaural Speech Enhancement (SE) in Hearing Aids (HAs) we propose an evaluation methodology. Instead of relying on test signals under lab conditions, we propose to leverage spatial audio for generating realistic speech-in-noise 3D test scenes, which are played back through a loudspeaker array and recorded with a KU100 dummy head, both wearing and not wearing HAs. The recordings are then used to compare the traditional HAs algorithms with offline DNN-based SE models, showing that the latter demonstrate higher consistency.

(ii) For normal-hearing monaural speech enhancement, we introduce the MB-RIRs dataset which examines three strategies for raising the ecological validity of synthetic Room Impulse Response (RIR) datasets — multiband absorption, source directivity and receiver directivity — and show that frequency-dependent absorption coefficients yield measurable gains when using real RIRs in the evaluation.

(iii) For Music Source Separation (MSS) of live recordings from the audience, where the performance of studio-trained models degrades because of ignoring the venue acoustics, the speaker system response and audience noise, we introduce CrowdioSet (real ambience tracks plus synthetic singalongs, generated from zero-shot singing voice conversions) and PaRIRset (stereo impulse responses captured across 40 real professional concert venues), and show that training on them improves separation under live conditions, both objectively and subjectively.

In each of the three tasks, we obtain performance gains without adding complexity at inference, investing it upstream in building more ecologically valid data instead.

Characterizing Error Distributions in Supervised Machine Learning

By: Abolfazl Mohammadi Seif
Supervisor: Dr. Ricardo Baeza-Yates

Date: December 21, 2026 - 11:00 h
Room: 55.309

Abstract

Predictive accuracy alone provides an incomplete picture of a machine learning model’s reliability, particularly in safety-critical applications where failures carry unequal costs. This thesis characterizes the distribution of errors across the model lifecycle, before training, during optimization, and at deployment, and turns that characterization into mechanisms that anticipate, estimate, and mitigate failures. Four contributions are presented. First, instance density and the Integrated Complexity Index (ICI) quantify data hardness before any model is trained. Second, a distribution-aware loss recovers bimodal error distributions and predicts classifier failure without accessing the deployed model. Third, Risk-Calibrated Learning and a backbone-agnostic post-hoc correction pipeline address the asymmetric cost of errors during training and deployment. Fourth, a learned trust score built on conformal prediction, Mahalanobis distance, calibrated confidence, and the predicted class determines whether an individual prediction should be accepted or referred to an operator. Extensive experiments on biomedical and vision benchmarks support these contributions, though evaluation remains retrospective rather than prospective.

Implementation of Artificial Intelligence-based Tools for Clinical Decision Making in Paediatric Cardiology

By: Adriana Modrego Muñoz
Supervisors: Dr. Patricia García Cañadilla, Dr. Joan Sánchez de Toledo & Dr. Bart Bijnens (UPF)

Date: January 26, 2027 - 12:00 h
Room: 55.309

Abstract

Postoperative management of Congenital Heart Disease remains a complex challenge due to high patient variability and the critical nature of recovery. This thesis presents the development of specialized Artificial Intelligence tools designed to support clinical decision-making in the Paediatric Cardiac Intensive Care Unit. By leveraging retrospective clinical data, the proposed models aim to provide accurate, longitudinal estimations of the Intensive Care Unit Length of Stay and dynamic assessment of patients’ clinical stability after surgery. The research focuses on identifying clinical indicators of stability, allowing for more proactive resource allocation and personalized care pathways. Results demonstrate that Artificial Intelligence - driven predictive tools can identify potential patient deterioration, offering a framework for clinicians to optimize discharge planning and improve patient outcomes. Ultimately, this work bridges the gap between complex data analysis and bedside application, enhancing the postoperative monitoring in paediatric cardiology. 

Optimal Feedback Control under Biologically Motivated Noise - When Internal Computation Is Unreliable

By: Francesco Damiani
Supervisor: Dr. Rubén Moreno

Date: February 4, 2027 - 11:00 h
Room: Auditori Mercè Rodoreda (Ciutadella Campus)

Abstract

Understanding how the sensorimotor system in humans and other animals sustains coordinated and adaptive behavior despite variability and uncertainty is a long-standing problem in computational and theoretical neuroscience, with important clinical and engineering implications. Normative models, particularly optimal feedback control, have long provided a principled approach for describing the computations underlying sensorimotor control. A widely used formulation within this framework is the Linear–Quadratic–Multiplicative– Internal (LQMI) problem, which extends classical Linear–Quadratic–Gaussian control with biologically motivated motor, sensory, and internal noise, all of which have proved important for explaining human motor behavior. Unlike sensory or motor variability, internal noise directly affects the internal representations and computations used to guide action, thereby reducing their reliability.

We first show that established solution methods for the LQMI problem rely on incorrect assumptions, leading to suboptimal solutions. We then derive an efficient analytical method to solve the problem in arbitrary dimensions. Using this framework, we investigate the role of internal variability, a source that has received comparatively less attention than sensory and motor noise despite its biological relevance. Our results indicate that internal variability induces a distinctive dynamical reconfiguration of the sensorimotor loop in which sensory information is weighted more strongly and its encoding in internal representations is reshaped, limiting the behavioral impact of internal fluctuations. Importantly, analogous signatures emerge from human tracking behavior when internal variables are inferred without imposing optimality. Finally, we extend the optimal-control problem introduced above by optimizing the internal dynamics themselves, relaxing the assumption that they must reproduce the dynamics of the physical world in the same state coordinates. This additional flexibility improves control under internal variability while preserving information about the physical state.

Overall, this thesis provides new analytical methods for stochastic optimal control under biologically motivated noise and identifies internal variability as a distinct computational constraint. More broadly, our results suggest that internal representations supporting robust behavior may depend not only on the statistics of the external world and on sensory and motor variability, but also on the reliability of the computations implementing them.