Chalmers Open Digital Repository
Not a member yet
26247 research outputs found
Sort by
Data driven heart failure patient segmentation: Identification of underlying patient phenotypes
No full textHeart failure is a heterogeneous syndrome where the underlying aetiology remains uncertain. The most common sub-classifications of heart failure are based on left ventricular ejection fraction, forming three subgroups: reduced (HFrEF), mildly reduced (HFmrEF), and preserved (HFpEF). While the development of targeted therapies holds promise for improving outcomes across the heart failure spectrum, current sub-classifications have not yet enabled precision medicine, as they fail to fully capture the underlying pathophysiological mechanisms.
Recent efforts have focused on HFpEF, and thus, patient segmentation in heart failure remains suboptimal. Based on data from four clinical trials in heart failure with a total of 11,140 patients, this project aimed to identify clinically relevant clusters across the full spectrum of ejection fraction. Twelve numerical and categorical patient characteristics were used as input covariates in latent class analysis to model underlying distributions, identifying four distinct clusters:
• Cluster 1: Old atrial fibrillation no myocardial infarction (HFpEF, 29.1% of patients)
• Cluster 2: Male high NT-ProBNP with myocardial infarction (HFrEF, 29.2% of patients)
• Cluster 3: Obese diabetic (HFpEF, 22.0% of patients)
• Cluster 4: Young male, with good kidney function (HFrEF, 19.8% of patients)
The identified patient sub-groups are clinically meaningful and associated with significantly different hard outcomes, such as cardiovascular and all-cause mortality. A sensitivity analysis using a k-prototypes clustering algorithm for mixed data derived clusters that corresponded to most characteristics of the HFrEF groups. However, cluster separation was relatively low across both latent class analysis and k-prototypes. Further work, such as data-driven feature selection, could improve cluster quality and separation
Reconstructing Private Data from Trained Models
No full textThis thesis investigates whether Model Inversion (MI) attacks can be effectively
adapted to tabular data—a domain where risks are underexplored compared to the
image modality. To address this question, we propose a novel adaptation of the
Pseudo-Label Guided Model Inversion (PLG-MI) attack for tabular data by utilizing
a Conditional Tabular Generative Adversarial Network (CTGAN). In support
of this contribution, new evaluation metrics are proposed—most notably, class-level
column shape scores—which serve to measure the similarity between reconstructed
and original private data. These metrics offer a practical means to evaluate the
privacy risks posed by inversion attacks in the tabular setting. As an initial step,
we have reproduced the PLG-MI attack for images, and verified that the attack is
robust even on target models trained on unbalanced private data sets. Then, by
applying the adapted tabular attack to a deep neural network diagnosis classifier
trained on the MIMIC-IV clinical dataset, we demonstrate that sensitive features
can be recovered with high accuracy. This shows that MI attacks can generalize to
the tabular domain; with default hyperparameters and minimal tuning, our method
recovers sensitive features with high accuracy. We have identified that the use of
transformations between structured and unstructured data, as well as the common
use of tree-based models in the tabular domain, can prevent adversarial gradient
access, thereby limiting the applicability of white-box model inversion attacks to
specific scenarios. Overall, our results confirm that model inversion attacks pose a
real privacy threat in the tabular domain while also clarifying the technical boundaries
that define when such attacks are viable. Our work will be made available as
a part of the LeakPro repository: github.com/aidotse/leakpr
Kinetic Simulations of the Proton-Alpha Instability in Collisionless Shocks
No full textThe solar wind is a dynamic stream of collisionless plasma originating from the
Sun. When it reaches the Earth’s magnetosphere, it abruptly slows down from super Alfvénic to sub-Alfvénic speeds, creating a bow shock. This compresses and heats
the inflowing plasma. The solar wind consists primarily of electrons, protons and
alpha-particles, which interact differently with the shock. A relative drift between
protons and alpha particles develops, which can trigger a streaming instability, driv ing waves. Such waves have previously been identified in spacecraft observations.
The dependence of the instability on the particle species temperature, density, and
flow velocity has previously been investigated using the dispersion relation for an
unmagnetized plasma.
In this work, we investigate this proton-alpha streaming instability using two dimensional fully kinetic Particle-In-Cell simulations. This approach offers the flex ibility to freely choose the initial conditions while simulating the full non-linear
dynamics. We study the parametric dependence of the instability on the flow ve locity of the alpha particles, the density ratio of the ions as well as the influence
of a background magnetic field. In particular, we want to quantify the heating of
different particle species due to the instability. Additionally, we aim to test the
applicability of the simplified linear theory.
Our simulations show that the instability mediates a strong energy exchange
from the flow speed of the alphas into plasma heating of the ions. We attribute
this to a Landau resonance between the proton and the alpha distribution, which
produces large plateaus in velocity space. The electron dynamics do not change
significantly. The wave properties are consistent with the linear theory when the
external magnetic field is small. Generally, the angle between the wave and the drift
direction increases with the flow speed. The fraction of energy converted to plasma
heating reaches a maximum at a wave angle of 45 deg to the drift direction. Overall,
the existence of a background magnetic field slightly suppresses the instability, due
to the magnetization of electrons. We conclude that the proton-alpha streaming
instability can play an important role in heating protons and alpha particles across
collisionless shocks
Increasing the use of battery electric trucks in line-haul operations An analysis of the enablers and barriers for electrification in the transportation system - from a network perspective
No full textThere is increasing pressure to decarbonize the transportation sector as a response
to the climate change, the climate targets, and the European Union´s target of
reaching net-zero emission by 2050. This Thesis investigates the introduction of
electric trucks (BETs), into DB Schenker´s line-haul operation, with focus on the
technical, operational, and economical feasibility. A mixed-method approach was
used, including a multiple case study, where three routes from DB Schenker´s trans-
portation network were analyzed, focusing on total cost of ownership (TCO), and
supported by interviews with key stakeholders involved in the transformation to
electricity on these routes. The Thesis also uses the ARA-framework to analyze
how actors, resources and activities influence and enable this transformation. The
findings show that while introduction of BETs offer both environmental and cost
benefits in the long run, the transition faces challenges in terms of infrastructure
development and cooperation between actors in the transportation ecosystem
Camera-based Vital Sign Detection in Autonomous Vehicles using Deep Learning - A Benchmark Study
No full textThis study explores the feasibility of employing camera-based, deep learning algorithms for detecting vital signs in autonomous vehicles, with a focus on enhancing driver safety. By evaluating various remote photoplethysmography techniques in dynamic vehicular environments, challenges such as motion artifacts and varying lighting conditions were addressed. Findings suggest that machine learning models, particularly neural network based approaches, hold promise in accurately estimating heart rate and respiratory rate from video data in such settings. The study emphasizes the potential of deep learning methodologies to improve driver safety through the integration of non-invasive vital sign monitoring systems in autonomous vehicles. Future research should address dataset imbalances and broaden the benchmark scope to include additional vital signs and algorithms, while also exploring alternative methods such as optical-flow based approaches to enhance respiration rate detection
Using Machine Learning to Develop Diagnostic Models for Cognitive Diseases
No full textVascular Cognitive Disease (VCD), including Subcortical Small Vessel Disease (SSVD), remains one of the most underdiagnosed cases of dementia. Given the clinical need for early, accurate, and explainable diagnosis, this study explores machine
learning techniques trained on real-world clinical data leveraged from the Gothenburg Mild Cognitive Impairment (MCI) study to uncover key variables in VCD diagnosis. The methodology was structured into three key parts: data preprocessing,
model training and evaluation, and model explainability. The original dataset was partitioned into three subsets to reflect distinct clinical settings: primary care, specialist care, and research data, each with varying levels of feature availability. Given the high rate of missing values, multiple imputation techniques were explored and assessed. The model training part involved evaluating the performance of various machine learning algorithms across specific diagnostic tasks and clinical settings. These machine learning algorithms included two gradient boosting tree algorithms XGBoost and LightGBM, Support Vector Machine (SVM), Random Forest (RF), and Naive Bayes. The model with the highest average F1 score (ranging from 0 to 1) across multiple iterations was selected for final deployment and further refined through additional training. An Explainable AI (XAI) approach, named SHapley Additive exPlanations (SHAP), was applied to the final model to ensure transparency and clinical relevance and identify the most influential features contributing to classification outcomes. The majority of the final models could classify diagnoses with high precision and recall, achieving high F1 scores. Some of the variables were previously known to be associated with the diseases. Furthermore, new variables not previously linked to the disease were identified, prompting further research. In conclusion, the machine learning pipeline built in this study has the potential to act as a classifier to distinguish VCD from AD and clinical prestages of cognitive impairment in a clinical setting. Furthermore, it can be utilized to identify key variables associated with distinguishing between these diseases
Microservice Granularity and Development Overhead
No full textMicroservices have become more prevalent recently, with several large enterprises adopting them. Microservices promise many advantages, but there are also challenges involved with deciding on their granularity, and this decision has many effects. The thesis explored through an investigative case study whether practitioners at one organisation perceived that a finer-grained microservice architecture resulted in additional work, and it was concluded that they did. Notably, they pointed out that intentionally suboptimal granularity decisions were frequently taken in the interest of time for services with rapidly changing requirements, and that this in part was due to the additional difficulty of splitting the service. Quantitative data from these practitioners’ organisation back this up. Other key results supported by both quantitative and qualitative data from the organisation are that the type of the microservice impacts how the granularity changes over time — that a small subset of microservices with rapidly changing requirements grow large, while most microservices tend to stay the same size after initial development work is done. It is then discussed how this growth can be seen as architectural technical debt caused partly by organisational processes, and that it would be beneficial for the organisation to consider this when planning. This adds further support to literature claiming that organisational processes are crucial for microservice development and that there are additional potential disadvantages beyond the immediately obvious to using microservices, and that it is crucial to consider carefully whether the advantages truly outweigh the disadvantages before choosing to use a microservice architecture, especially for smaller systems
Developing a Back-end Compatibility Check for the BNF Converter
No full textThere exist various architectural models for structuring large-scale software projects. A common pattern employed in various systems involves decomposing the application into a user-facing front-end component, responsible for input validation, and several back-end components, which perform subsequent processing and generate output. In such a system, each back-end may target a distinct output format. Communication between the front-end and the back-ends is typically conducted by means of a well-defined interface.
It is often desired that the front-end be responsible for all input validation, so as to promote encapsulation and enforce uniform communication to the user. However, difficulties may arise when the back-ends impose different requirements on the input that they process.
This thesis examines this problem in the context of the BNF Converter – a compiler construction tool that transforms a common grammar specification into software components used for implementing a compiler for the language of said grammar. The BNF Converter adheres to the aforementioned architectural model and supports multiple back-ends that target various output formats and are characterized by distinct capabilities and requirements with respect to the input.
In order to address these discrepancies, a so-called back-end compatibility check was developed. This check enables the front-end to validate whether a given input conforms to the requirements imposed by a particular back-end. The methodology used for this purpose is expected to generalize to other systems exhibiting a similar structure. Additionally, a new back-end was implemented, both to evaluate the usefulness of the interface and to contribute to the development of the BNF Converter itself
Användarcentrerad utveckling av en vårdapp för kvinnor
No full textThis project presents the design and development of a digital healthcare application tailored to women’s needs across all various life stages, from menstrual health and contraception to pregnancy and menopause. The goal was to design a reliable and user-friendly interface that addresses common issues in Femcare’s current digital application. Such as poor personalization, unclear information, and limited usability on digital platforms. The project followed a user-centered design process starting with research and evaluation of the existing mobile application. This was followed by prototyping a new interface and testing its usability. The results indicated improvements in navigation clarity, task completion time and user satisfaction. A key design challenge was striking a balance between a personal tone and the need of digital privacy, to ensure the approach remained appropriate and relevant for a healthcare application. The final prototype reflects an intuitive and inclusive design with the potential to complement and improve the existing digital healthcare services for women