Abstract:
Aim:
This project investigates the unexplored clinical potentials of photoplethysmography (PPG) as an assessment tool for patients with atrial fibrillation (AF). We aim to investigate (1) the impact of risk factors on PPG, (2) how AF-related hemodynamic changes are reflected in PPG, and
(3) how ablation treatment for AF affects hemodynamics.

Background:
PPG is a technique that uses light to detect volumetric changes in the peripheral vasculature. It is widely available in wearables and provides a more continuous signal than electrocardiography (ECG). In research, PPG has been used to detect AF with high accuracy comparable to ECG. PPG is less well characterized than ECG, and it is unknown how ageing, hypertension, diabetes, and other risk factors as well as hemodynamics relating to AF are reflected in the PPG. This project will generate important basic knowledge on the clinical use of PPG and at the same time investigate the hemodynamics of AF, the most common arrhythmia worldwide.

Methods:
We will develop deep neural networks (DNN) for detecting hemodynamical patterns related to AF based on PPG recordings and characteristics from three independent cohorts comprising >6500 patients. Specifically, we will apply a DNN to (1) use PPG signals to distinguish patients with a risk factor (e.g. diabetes) from a patient without, (2) investigate how the hemodynamic changes before, during and after AF in PPG signals, and (3) distinguish between a patient’s hemodynamical pattern before and after they have received ablation therapy with PPG signals. To allow for linkage between the PPG signal and the outcome, we will specifically develop and apply explainable AI (xAI) methods for PPG analysis. xAI allows for a visual interpretation ofthe otherwise hidden decision-making of the DNN and graphically depicts the linkage of the
signal to the outcome. xAI has previously been used with ECG analysis and in this project, we will develop the method for use with PPG signals for characterisation of hemodynamics associated with the risk factors, paroxysmal AF, and AF management.

Perspectives

This project will provide a novel understanding of PPG necessary for future clinical use and investigate unknown mechanisms of AF. Firstly, we will characterize the effect of prevalent risk factors on the PPG with huge implications for PPG algorithm development. We will also determine to what degree PPG may be used as a gatekeeper for further diagnostic work-up and reduce the number of unnecessary tests for the benefit of patients and society. Secondly, we will generate important knowledge on AF mechanisms and on how hemodynamics are reflected in PPG signals and our findings will be part of the scientific foundation necessary for the use of PPG in healthcare, whether driven by industry or academia. Finally, this project will help gain mechanistic information on ablation as a treatment for AF and might eventually help inform personalized treatment.

Abstract:

Automac Anomaly Detecon in Health Registry Data by Dynamic, Unsupervised Time Series Clustering

Denmark has established a wealth of health registries used to monitor the quality of health care. Although this resource has enormous potential, data has become so complex and highdimensional that important insights in quality of care and patients’ safety may go unnoticed. There is, thus, a need for a dynamic, automated algorithm capable of flagging growing anomalies in registry data, helping health care personnel to rapidly discover important divergencies.

In this project we will develop and test a new algorithm based on dynamic, unsupervised time series clustering with anomaly detection for health care data. At each time point, the algorithm will cluster patients (using, e.g., hierarchical, t-SNE, or autoencoder clustering) based on a patient trajectory metric (e.g., Hamming distance or optimal matching) and the development of anomaly clusters will be monitored by significant change in a cluster dissimilarity measure (e.g., Jaccard distance or MONIC). Thisalgorithm’s output will consist of summaries of detected anomalies in a form that allows for a quick assessment by relevant health care professionals. These summaries will be evaluated by a team of experts, and the algorithm will be tuned based on their input. The algorithm will thus learn, through supervision, to predict expert interests. The algorithm will be developed and tested on the Danish Diabetes Database (DDiD).

Such an algorithm would greatly improve the health care system’s ability to react timely on both positive and negative trends in quality of care. Furthermore, the algorithm will be developed in a disease independent fashion, such that it can be implemented more generally and potentially be used to monitor other areas in critical need of attention.

Abstract:

More than 650 million people suffer from prediabetes worldwide and the prevalence is increasing rapidly. A large part of these people will eventually develop microvascular and macrovascular complications generating a large economic burden on society. To prevent or delay onset of these complications, both lifestyle and pharmacological interventions are necessary. However, treatment tools or guidelines specifically for prevention of complications for this group does not exist in general practice. To address this challenge, the project seeks to improve the management of people with prediabetes by developing a decision support system to be implemented at the general practitioner. Based on a prediction of the personalized risk of micro- or macrovascular complications and a risk stratification, individuals with high- risk profiles will be identified. Additionally, different scenarios with lifestyle and pharmacological interventions will be simulated. This novel prediabetes risk engine tool will support informed treatment and early prevention strategies at the general practitioner, aiming to prevent or delay the onset of complications. Data from clinical studies and Danish national registers will be analyzed using data science techniques to identify patterns which are important for prediction of diabetes-related complications. Additionally, Artificial Intelligence including machine learning methodology will be used to develop the prediction model. No studies regarding prediabetes have investigated development and implementation of a flexible model allowing usage with only a limited amount of clinical data, and with a possibility of entering further data to increase precision of the risk estimate. Therefore, this project will focus on development of a flexible predictive model aimed at estimating the personalized risk of micro- or macrovascular complications among individuals with pre-diabetes.

Abstract:

Abstract:

Cardiometabolic diseases including type 2 diabetes (T2D), cardiovascular disease (CVD), and obesity pose a growing global health problem, and a decay in the public metabolic health in Greenland is associated with westernization of diet and lifestyle. The genetic architecture of the modern Greenlandic population is shaped by its demographic history, geographic isolation in an Arctic
climate, and small population size, resulting in strong genetic drift and a high frequency of highimpact gene variants. Although genetic variants with high impact on metabolic health have already
been described, the genetic regulation of the plasma lipidome and its link to cardiometabolic diseases is poorly understood.

Using a state-of-the-art high-throughput mass spectrometry-based lipidomics, we aim to integrate plasma lipidomics data and genetic data from 2,539 Greenlandic individuals to better understand the link between lipid species and metabolic health. A study visit to Swedish University of Agricultural Sciences for collaboration will provide this project with nuclear magnetic resonance lipidomics
analysis, contributing with quantitative and qualitative knowledge of the study population and the potential identification of novel compounds. With genome-wide association studies, mapping of lipid quantitative trait loci (lQTLs), and colocalization analyses, we will examine the cross-sectional associations between lipid profiles, registry-based data on cardiometabolic outcomes, and genetic data to identify prognostic biomarkers and investigate biological pathways related to T2D, CVD, and
obesity. We hypothesize to see changes in the plasma lipidome in genetic loci linked to cardiometabolic disorders due to genetic drift of the Greenlandic population accompanied by westernized diet and lifestyle.

This project could offer novel insight into genetic etiology of cardiometabolic diseases to improve our understanding of molecular disease mechanisms and reveal novel targets for disease treatment and prevention in a broader perspective. The discovery of novel high-impact genetic variations associated with altered lipid profiles can contribute to understanding of metabolic health in Greenland and highlights the implications this research has for genetic precision medicine

Abstract:

Abstract:

Diabetes and prediabetes are increasingly prevalent conditions in modern society, both of which are associated with numerous health hazardous conditions such as obesity, hypertension, and cardiovascular disease. In itself, type 1 diabetes (T1D) is a life changing diagnosis, forcing a need for constant health awareness. When dealing with these challenges, a continuous glucose monitor (CGM) is a vital tool that helps patients evaluate their own health and helps inform clinical decision making in a cost effective manner. Use of CGM devices is therefore becoming more and more common in diabetes clinics around the world, where data from CGMs are collected and analyzed with the objective of optimizing patient care. The increasing adoption of CGMs brings about a huge potential for improving care by developing a data-driven methodology that can be used to assess the CGM data. However, since only simplistic methods based on different summary statistics have been attempted in clinical practice, we still need to uncover the full potential of the information production in CGM measurements.

Abstract:

Abstract:

The ability to predict disease risk and identify individuals at high risk for developing a certain disease is fundamental to modern healthcare, since it enables the implementation of preventive measures and personalized treatments. Polygenic scores (PGS) have received attention for their promise to improve clinical prediction models. Recently, electronic health records (EHR) have also proven to enhance prediction accuracy. However, the accuracy of both PGS and EHR in clinical prediction models is impacted by individual genetic, environmental and diagnostic heterogeneity, which can lead to racial, gender, and ancestry-based biases. It is important to understand and measure the impact and severity of these types of heterogeneities, in order to develop more inclusive, accurate and robust prediction models. These models need to be evaluated and replicated across cohorts and in individuals of different genetic ancestries.

Abstract:

Two-thirds of human hormones act through ~800 G protein-coupled receptors (GPCRs). The vast majority (71%) of these hormones are peptides or proteins, which also account for an increasing share of drugs. The study of peptide-receptor recognition is thus essential for understanding physiology, pathology and for drug design.

This project aims to solve the modeling problem of peptide-receptor recognition by leveraging machine learning methods and unique data from the field hub GPCRdb. I will build predictive graph neural network models representing residue interaction networks across the receptor-peptide interface. The models will utilize attention-based transformer and LSTM architectures which have shown great promise in drug-target interaction prediction and de novo-drug design. The models will be trained on a unique data representation, storing data for individual residues rather than the overall protein. This will allow peptide data to be inferred across conserved residues in different receptors – enabling use on receptors not targetable with classical methods.

The trained models will be used in three applied aims to: (1) discover peptide hormones by matching the library of predicted physiological peptide hormones to their cognate receptors with unknown physiological ligand and function; (2) identify peptide probes by matching pentameric peptide library to understudied and drug target receptors; and (3) holistically engineer probes for those receptors residue-by-residue. The in silico discovered probes will be tested in vitro by pharmacological collaborators. In all, this will let me discover novel hormones and engineer new probes, enabling functional characterization of understudied receptors that cannot be targeted with current techniques.

This project has the potential to uncover mechanisms of peptide-receptor recognition underlying physiological, sensory, and therapeutic responses. This will lay the foundation for exploring uncharted receptor functions and designing better drugs. Given this and that our approach will be applicable

Abstract:

Abstract:

Abstract:

Abstract:

This research proposal is a collaboration between the Machine Learning Group at University of California, Berkeley, Bruce Fischl’s group at Harvard Medical School, and the Medical Image Analysis Group at University of Copenhagen.

Deep learning has shown tremendous success in brain image analysis, aiding clinicians and researchers in detecting and treating a wide range of diseases from various data sources such as CT and MR images. A common task is to perform segmentation on images. However, the field is fundamentally limited by a lack of labeled training data and substantial variations in the data available. As a result of this, models often exhibit a lack of robustness to changes in equipment, patient cohorts, hospitals, and scanning protocols.

Because segmentation models have very large hypothesis spaces, existing domain adaptation theory and methodologies have failed to alleviate these problems. Current methodologies are either not grounded in theory or impractical to apply to segmentation models, where the large hypothesis spaces make it intrinsically difficult to overcome numerical issues or achieve noteworthy performance.

To push forward the field of brain image analysis, there is a need for theoretically well-founded domain adaptation methods. This project aims to work towards such methods, by boldly conducting theoretical work with the world-leading machine learning group at Berkeley, and apply this work to brain image segmentation, in collaboration with Bruce Fischl’s world-leading group at Harvard Medical School. The project is to be centered at the Medical Image Analysis group at University of Copenhagen, which is internationally recognized for applying cutting-edge machine learning practices to medical image analysis problems.

If successful, this project will lead to more robust models, a fundamental contribution towards utilizing the vast amount of medical images which is being produced at hospitals worldwide. This work will contribute towards providing technology for improving fundamental brain research as well as.

Abstract:

Abstract:

Abstract:

Abstract:

Abstract:

Abstract:

Abstract:

Misinformation and propaganda are recognised as a major threat to people’s judgement and informed decision making in health, politics and news consumption. The spread of misinformation relating to the Covid19 epidemic is just one prominent example. It is not only wrong facts that constitute a threat, but also the language used which can lead to deception and misleading of people. To address the misinformation threat and empower readers confronted with enormous amounts of information, we propose a new data science methodology for the computational analysis of rhetoric in text.

Our brain is the most central part of the nervous system and crucial to our health and wellbeing. In Denmark, mental disorders make 25% of the total disease burden, with yet increasing prevalence. Across many medical domains, magnetic resonance imaging (MRI) is a widely used tool to study the anatomy and physiology of soft tissue, e.g. the brain, and is applied in diagnosis and monitoring of diseases, as well as a tool to investigate their underlying mechanisms. In psychiatry, research has yielded substantial evidence for structural brain changes at a group level, however these are typically subtle and currently, there is no clinical benefit from MRI for the individual patient.Previous research aiming to identify brain aberrations in patients with neuropsychiatric disorders struggle with relatively small and often inhomogeneous samples paired with complex clinical traits and weak pathological signals.

To unravel the intricacy of mental disorders and the brain, one approach is to apply powerful state-of-the-art machine learning algorithms, such as deep neural networks (DNNs). Large DNNs are able to extract high level features of images and other signals and able solve sophisticated tasks, outperforming traditional machine learning methods by far. On the downside, a DNN requires a large amount of ‘labelled data’, e.g., where each brain image has a meaningful notation, such as ‘patient’ or ‘control’. The amount of labelled data needed to train such a model is currently not available in conventional psychiatric research. In contrast, ‘unlabelled data’, e.g. normative brain images independent of a certain class or group, are often generously and publicly available. In the field of machine learning, scarcity of labelled data and richness of unlabeled data has given rise to self-supervised learning paradigms. In self-supervised learning one exploits rich unlabeled data to learn a general intermediate representation of the matter of interest. Scarcelabelled data is used efficiently to fine-tune the intermediate representation to a specific task of interest.

The aim of this project is to develop a self-supervised DNN model of the brain using MRI data of large international, high-quality databases. This model will then be fine-tuned using highly specific data from psychiatry to address specific research questions regarding the mechanisms of aberrant neurodevelopment. We believe a self-supervised model is more robust and able to learn more meaningful features compared to conventional models. To explore these features and their relation to clinical traits present in psychiatric patients, we want to employ explainable artificial intelligence techniques.

To sum up, we want to use self-supervised learning paradigms and utilize its efficient use of scarce labelled data to develop a state-of-the art DNN model of brain images, bringing neuropsychiatric research to the forefront of machine learning research.

Note: Since the date of recording the video, Fabian has chosen to adjust the scope and title of his research project, the title in the written article is current and correct

Somatic mutations play an integral role in the development of cancer. In the past decade the identification of patterns in the somatic mutations, called mutational signatures, has in- creased in popularity. These signatures are associated with mutagenic processes, such as DNA damage and sun exposure. Although the signatures contain vital information about tu- morigenesis, there is a lack of confidence in the signatures which are estimated predomi- nantly by non-negative matrix factorisation.

We propose an autoencoder alternative to sig- nature extraction which we hypothesize will increase stability and confidence in the signa- tures. These new signatures will be used to diagnose ovarian cancer patients with homolo- gous recombination deficiency, a DNA deficiency that has been shown to be sensitive to PARP inhibitor treatment. Potentially, this test leads to improved identification of ovarian cancer patients who will respond to platinum treatment, a surrogate treatment for PARP inhibitors, which would indicate that the proposed test could successfully act as a predictive biomarker for PARP inhibitor treatment.

The project will deliver a pipeline for confident stratification of cancers based on mutational signatures, providing one step further towards personalised medicine for DNA repair-defi- cient tumours.