Theses Titles and Abstracts

Austin Drake Functional vs. Structural Brain Measures: A Comparison in Predicting Physical Activity and Aerobic Fitness Austin Drake This work was carried out under the supervision of Professors Tiina Parviainen and Jan Kujala, Faculty of Education and Psychology, University of Jyväskylä. Interdisciplinary Studies Unit, The Leslie and Susan Gonda Multidisciplinary Brain Research Center, Bar-Ilan University. ABSTRACT Recent studies have shown interactions of physical activity and aerobic fit ness with various brain features and neural dynamics. However, while physical activity and aerobic fitness are distinct concepts, differences in their unique in teractions with the brain are unclear. There is existing evidence that structural features such as gray matter volume correlate with aerobic fitness but not physi cal activity, while functional features like MEG power spectral asymmetry corre late with physical activity and not aerobic fitness. The aim of this thesis was to determine if the associations of fitness with structural measures and physical ac tivity with functional measures are generalizable. To this end, multivariate non linear regression was used to test for widescale, complex effects. Cortical gray matter thickness and subcortical gray matter volume were used as the structural features, and resting state MEG power spectra were used as the functional fea tures. Additionally, MEG alpha frequency band power was tested separately, as well as a combination of gray matter thickness/volume and MEG alpha power. SVM regression models with linear and RBF kernels were trained to predict av erage moderate-to-vigorous physical activity (MVPA) & log-transformed MVPA levels, as well as 20-m Shuttle run times using Leave-One-Out cross validation. Permutation testing was used to validate the performance of each model. None of the models were found to predict either MVPA, log-transformed MVPA or 20 m Shuttle run times above chance level. This suggests that multivariate non-lin ear regression analysis at the present scale may be insensitive for predicting spe cific fitness levels, or that the analysis requires further refinement to observe the anticipated effects.

Kanykei Mairambekova Classifying Aerobic Fitness and Physical Activity Levels from Structural and Functional Brain Connectivity Data: A Multimodal DTI and MEG Machine Learning Study This work was carried out under the supervision of Prof Tiina Parviainen and Jan Kujala, Faculty of Education and Psychology, University of Jyväskylä. Interdisciplinary Studies Unit, The Leslie and Susan Gonda Multidisciplinary Brain Research Center, Bar-Ilan University. ABSTRACT The relationship between physical activity, aerobic fitness, and brain connectivity has not been studied thoroughly in adolescents, especially using com-plex data driven approaches. This thesis aims to examine whether the high and low levels of physical activity and aerobic fitness in adolescents can be classified based on structural and functional connectivity measures using supervised machine learn ing models. Structural connectivity was derived from DTI data and included av eraged fractional anisotropy and mean diffusivity values across tracts. Functional connectivity was assessed using resting state MEG (rsMEG) coherence estimates across six frequency bands. The classification models based on structural connectivity achieved performance above chance level: 64% balanced accuracy for aerobic fitness and 61% balanced accuracy for physical activity. In contrast, functional connectivity models using rsMEG coherence, even after applying several dimensionality reduction tech niques such as PCA, frequency band specific models, and feature preselection, did not outperform the chance levels in either of the classification tasks: around 50% balanced accuracy was reached for both tasks. These findings show that structural connectivity features can be used to classify physical activity and aerobic fitness levels in adolescents, whereas resting-state MEG coherence did not support accurate classification in the current study. Moreover, the results demonstrate that machine learning can help to find com plex patterns in brain connectivity data and give insights when conventional ap proaches yield limited results. Keywords: physical activity, aerobic fitness, structural connectivity, functional connectivity, machine learning approach

Aracely Gutiérrez Lomelí Brain-Body Coupling: Investigating Heart Rate Variability States with Magnetoencephalography across Respiratory Patterns This work was carried out under the supervision of Professors Tiina Parviainen and Jan Kujala, Faculty of Education and Psychology, University of Jyväskylä. Interdisciplinary Studies Unit, The Leslie and Susan Gonda Multidisciplinary Brain Research Center, Bar-Ilan University. ABSTRACT Unravelling the riddles of the constant interplay between modulatory pathways between the brain and the body is important for various applications in research and clinical fields. Unveiling how information related to autonomic regulation is encoded in the brain can be translated into a better understanding of emotional and physiological traits by investigating the recorded neural signals. Previous studies show heart and respiratory coupling with neural spectral patterns, suggesting heart rate variability (HRV) can be decoded from brain sig nals. This study aimed at decoding HRV states, particularly its low frequency (LF) component, using the spectral content in neural signals recorded with mag netoencephalography (MEG) during so called resting state. The data collection involved 33 participants, with electrocardiogram (ECG) signals recorded simul taneously with MEG. This study's novelty lies in a data-driven segmentation ap proach to epoch MEG and ECG signals according to natural HRV fluctuations. Eighteen variables were assessed in a logistic regression machine learning (ML) model, where each MEG segment was a sample, and model features con sisted of MEG power across five frequency bands (delta, theta, alpha, beta, and gamma). The variables included three respiratory patterns (Deep breathing with eyes open, spontaneous breathing with eyes closed, spontaneous breathing with eyes open) × three MEG sensor types (magnetometers, gradiometers, or both) × two ECG segmentation methods (increasing and decreasing HRV gradient trends, or high and low HRV segments). The decodability of HRV variability from MEG data could not be confirmed since models performed at chance level. Future research suggestions are provided to improve the results.

Arna Aimysheva Decoding Arousal and Valence from Continuous MEG Data during Video Watching with Machine Learning This work was carried out under the supervision of Prof. Tiina Parviainen and Jan Kujala Faculty of Education and Psychology, University of Jyväskylä. Interdisciplinary Studies Unit, The Leslie and Susan Gonda Multidisciplinary Brain Research Center, Bar-llan University. ABSTRACT Mental disorders represent a prevalent and rising global health condition that is often closely associated with certain emotional states. Identification of emotional states based on brain function would have important consequences for mental illness diagnosis and personalized treatments. The current thesis investigates the categorization of emotional states and specifically arousal and valence levels based on the XGBoost machine learning model used on frequency-domain MEG data. Emotion-inducing one minute video clips were shown to participants while MEG signals were recorded. After each video the participants evaluated their arousal and valence levels, which were later binarized based on the mean arousal and valence ratings. The XGBoost classifier achieved 0.86 and 0.78 classification accuracies for valence and arousal, respectively. Feature importance analysis agreed with previous findings, showing the importance of beta band activity with respect to arousal and alpha band activity with respect to valence. These results contribute to the understanding of emotion decoding from brain activity and demonstrate the potential of machine learning techniques in affective neuroscience. This combination provides potential applications for improving mental health diagnostics and therapeutic strategies.

Nele Felicitas Werner A Convolutional-based Deep Learning Model Detects Neural Modulations Following Chronic μECoG-based Cortical Stimulation This work was carried out under the supervision of Prof. Dr. Klaus Linkenkaer Hansen, Faculty of Science and Vrije University Medical Center Vrije Universiteit. Abstract Brain stimulation targeting the mesoscopic level is increasingly used to treat neurological disorders by inducing excitability changes at specific brain regions, yet the neural e ects after chronic stimulation remain incompletely understood. This study investigated whether chronic cortical stimulation induces measurable changes in brain dynamics by comparing resting-state activity recorded using micro-electrocorticography ($\mu$ECoG) before and after stimulation in a sheep model (\textit{n} $=$ 8, 55 recording sessions each). Specifically, it was tested whether such changes can be detected using deep learning, as its ability to process neural data near real-time makes it a promising tool for future time sensitive closed-loop neuromodulation systems. To this end, both feature extraction (relative power, detrended fluctuation analysis and spectral entropy) and three convolutional neural network-based models (EEGNet, CAM-EEGNet, and CAM-EEGNet-TAM) were applied. Features di erences were significant, though only weak and inconsistent. Among the deep learning models, only EEGNet consistently classified pre- and post-stimulation above chance level with an accuracy of 71\% on average. The other two models performed inconsistently and failed to generalize across animals. These findings indicate that EEGNet was able to detect stimulation-induced changes in neural dynamics within individual animals, more e ectively than the chosen feature analysis or the two other deep learning models. However, performance varied across animals, indicating that the stimulation e ects were subtle and not consistently detectable. Further work is needed to better understand the underlying neural mechanisms to develop more interpretable models that can support future closed-loop neurostimulation applications.

Giulio Benedetti Exploring Inhibitory Control and Attention in Childhood with Neural Decoding University of Jyväskylä. Faculty of Education and Psychology ABSTRACT Inhibitory control and attention represent central mechanisms of cognitive functioning that are continuously recruited during daily activities. Their impairment is linked to various neurodevelopmental conditions, including Autism Spectrum Disorder and Attention-Deficit/Hyperactivity Disorder. As early cognitive development faces new challenges due to the widespread adoption of modern technology, it has become increasingly relevant to study the neural basis of inhibitory control and attention during childhood. Here, we investigated whether inhibition is separable from I) passive and II) attention conditions based on neural activity in the frequency domain. Additionally, we aimed to identify features that significantly contributed to this objective. To address these questions, we analysed patterns of neural activity associated with inhibitory control and attention in a cohort of 67 school-aged children using a neural decoding approach. During a magnetoencephalography experiment, participants performed go/no-go and oddball tasks to probe cognitive processes related to inhibition and attention. A decoder was then trained on frequency profiles of neural activity that emerged upon execution of the two tasks. Finally, contributions of distinct neural oscillations were examined via model performance and feature importance analysis. Results show that inhibition and attention can be accurately decoded from neural oscillations. Moreover, activity in higher frequency bands, especially gamma, primarily contributed to decoding performance in both cases. Overall, we report evidence that inhibitory control and attention partly rely on shared neural resources. By combining biological and computational expertise, we propose a knowledge-driven decoding approach that emphasises the strengths of integrating the fields of brain and data science.

Nghi Nguyen Amplified Cortical Excitation-Inhibition Biases Drive Static and Dynamic Functional Alterations in Major Depressive Disorder This work was carried out under the supervision of Prof. Martijn van den Heuvel, Department of Complex Trait Genetics, Center for Neurogenomics and Cognitive Research, Vrije Universiteit Amsterdam. Abstract As Major Depressive Disorder (MDD) is increasingly understood as a systems-level disorder, establishing a mechanistic link between its characteristic macroscale functional alterations and underlying neurobiology is becoming a key area of investigation. Resting-state functional neuroimaging studies have identified two key signatures in MDD: less well-defined functional boundaries across transmodal networks, and a brain-wide reduction in the turbulence of BOLD fMRI signals. We tested the hypothesis that these alterations can be mechanistically explained by a dyregulation of the transmodal networks’ excitation-inhibition (E-I) balance. Through a dual investigation combining case-control resting-state fMRI analysis (n=430) with in silico experiments using a heterogeneously parameterized neural mass network, we demonstrated that a uniform global shift in E-I balance failed to sufficiently and consistently capture these signatures found in MDD patients. Instead, our model suggests that they consistently emerge from an amplification of the pre-existing spatial gradient of E-I bias, whereby highly-excited regions become even more excited and vice versa. This “rich-get-richer” mechanism provides the first model-based evidence linking the static and dynamic signatures of MDD to a specific pattern of E-I dysregulation, implying that the disorder may be caused by a disruption in the brain’s homeostatic plasticity

Irene Bernardi Ultradian Heart Rate Variability as a Marker of Sleep Quality: A Wavelet-Based Analysis of Wearable Data in Adults at Risk for Cognitive Decline On-Site Supervisors: Dr. Ruud van Stiphout, IMEC The Netherlands Alex van Kraaij, M.Sc., IMEC The Netherlands Abstract Heart rate variability (HRV) is intimately linked to autonomic function, cognitive performance and sleep. Ultradian HRV rhythms may provide unique insight into nocturnal autonomic function, but have not been explored in older adults at risk for cognitive decline or in the context of lifestyle intervention. This exploratory study observed ultradian HRV in 63 participants (aged 60-75) at risk for cognitive decline, before and after a multidomain lifestyle intervention aimed at improving their health. Wavelet ridge periods were extracted from nightly recordings of a smartwatch in the 60-150 minutes range, i.e., the physiological range of sleep. Prior to the intervention, slower ultradian HRV rhythms were significantly associated with poor sleep quality. This association was not found in nightly average HRV. After the intervention, the low- and high-intensity groups showed differing ultradian HRV—sleep quality relationships, but neither reached significance. Overall, the intervention did not generate significant changes in ultradian HRV or sleep quality. After the intervention, all analyses were performed on small sample sizes, which limited statistical power and likely contributed to the lack of significant effects. Before the intervention, ultradian HRV emerged as a potential marker of sleep quality in older adults at risk for cognitive decline. To determine the potential of lifestyle interventions, future work should examine larger cohorts. Further, our analysis pipeline should be tested on gold-standard HRV recordings and validated on healthy individuals.

Cristian Douglas Sales Garcia Detection Optimization and Analysis of Absence Seizures in Physiological Signals of the Genetic Absence Epilepsy Rats from Strasbourg This work was carried out under the supervision of Prof. Sandra Cristina Henriques Vaz, Institute of Pharmacology and Neurosciences, University of Lisbon. Abstract Absence seizures (ASs) are non-convulsive seizures that are characterized by a brief loss of consciousness. These seizures typically occur in patients with absence epilepsy and are often present with distinctive 2.5-4 Hz spike-wave discharges (SWDs) in the electroencephalogram (EEG). In the Genetic Absence Epilepsy Rats from Strasbourg (GAERS), there is a noticeable behavioral arrest in synchrony with 5-12 Hz SWDs. No single automatic SWD detection algorithm has been met with consensus in the field, and currently used semi-automatic approaches are time-inefficient and makes room for experimenter’s bias. Here, we present two studies separately tackling this experimental constraint. First, using Python v3.11.11, we automatized the SWD detection using a combination of parameters from well-established features of the GAERS EEG. The detection algorithm presented here is optimizable, making it adaptable to different laboratory environments, and couples video recordings to facilitate assessment of synchronous behavioral arrest. As a validation, we analyzed 50 one-hour EEG recordings from a pharmacological experiment using GAERS treated with ethosuximide and valproate. Second, we tested the addition of accelerometer and load cell data in the current experimental setup, recording one Wistar rat recovering from isoflurane induced anesthesia. We show a validated detection algorithm and positive perspectives towards novel methodological enhancement in biosignal acquisition in the GAERS. These improvements may contribute to pre-clinical future studies of drug discovery to enhance treatment outcomes in patients with ASs.

Kseniia Nikulina Polynomial Image Formation and GAN-Based Decomposition for Poisson–Gaussian Noise Modeling This work was carried out under the supervision of Prof. João Miguel Raposo Sanches and Eng. Diogo João Fróis Vieira, Bio-Engineering faculty, Universidade de Lisboa. Abstract Poisson noise remains a major obstacle in biomedical imaging, as it introduces uncertainty, conceals anatomical structures, and cannot be treated as purely additive or multiplicative due to its signaldependent nature. While numerous denoising methods attempt to suppress this noise, they typically ignore the underlying noise field, which may itself hold diagnostic value. This thesis proposes a hybrid deep-learning framework that reconstructs the noisy observation through a pixel-wise polynomial image formation model. Designed to approximate the physical acquisition process under photon-limited conditions, the polynomial combines signal–noise interactions via real-valued coefficients, enabling a realistic representation of Poisson–Gaussian corruption. The architecture employs two Generative Adversarial Networks trained sequentially but guided by a unified objective: the first GAN estimates the clean image, and the second uses original noisy-clean pair of images to infer and model the corresponding noise field. The reconstructed noisy image is then synthesized from the denoised image and the estimated noise via the polynomial model. Training and evaluation were conducted on the Fluorescence Microscopy Denoising dataset using tailored priors and loss functions consistent with the formation model. Results demonstrate stable convergence, accurate noise-field recovery, and competitive performance against state-of-the-art baselines, underscoring the potential of explicit noise modeling in biomedical imaging workflows.