We have one PhD vacancy within the Centre. Details are:
Advanced epileptic seizure warning methods
Project Leader: David Grayden
Staff: Anthony Burkitt, Mark Cook
Primary Contact: David Grayden (firstname.lastname@example.org)
Keywords: biosignals; computational neuroscience; electroencephalogram EEG; epilepsy; neuroengineering
Disciplines: Biomedical Engineering
Domains: Convergence of engineering and IT with the life sciences
Research Centre: Neuroengineering Research Laboratory
This project will develop epileptic seizure prediction methods, warning patients of the likelihood of an impending seizure, so that precautionary measures can be taken. Seizure prediction will be of great clinical significance as it will improve the lives of 33% of epileptic patients who have drug-resistant epilepsy, by warning of impending seizures and potentially allowing acute therapies to prevent seizures, such as electrical stimulation of the brain or targeted drug delivery.
Description of the project environment:
There are a number of projects embedded within the Centre, where we have 14 PhD students, 5 post-doctoral researchers, and 11 University of Melbourne staff collaborating across our 4 research streams with researchers at IBM Australia and a number of other Australian Universities, including RMIT University, the University of South Australia, and Western Sydney University. Each student is jointly supervised by Universities and IBM Research and will spend at least 12 months of their candidature working on site at an IBM Research laboratory.
Our PhD students are working on the following projects:
Literature-based Discovery for Alzheimer’s Disease
This project will focus on methods for literature-based discovery of insights from the biomedical literature, with a particular focus on Alzheimer’s Disease. This will include both information extraction of elements related to the molecular basis and clinical phenotypes of the disease, as well analysis of the certainty of that information and inference to support assessment of consistency between information extracted from various sources.
This project will also benefit from the context of the Natural Language Processing group at the University of Melbourne, including Professors Tim Baldwin and Trevor Cohn, and other staff including Dr Jey-Han Lau, Dr Daniel Beck, and Dr Lea Frermann, and their existing students and research fellows, and Professor Karin Verspoor from RMIT University. Professor Baldwin is the Director of the training centre, while Professor Verspoor is the Deputy Director.
Using modern Machine Learning for probabilistic modelling of dementia
The central goal of this project is to develop novel predictive tools that use multi-modal data (demographic, clinical, imaging and genetics) to statistically model age of onset of Alzheimer’s Disease (diagnosis) as well as its subsequent trajectory (prognosis). These tasks come with a non-standard set of challenges (e.g. large chunks of missing data, dataset shifts across cohorts, frequent mislabelling, undocumented comorbidities and sparsity of observations compared to the disease timescale), thus representing a relatively uncharted territory for modern AI. On the quest to find the right tool for the job, you are likely to explore the entire spectrum of Machine Learning: from SVMs to Gaussian Processes, from PCA to Variational Autoencoders, from logistic regression to GBMs, from linear models to Deep Learning and so on.
Key areas involved in the research: machine learning, Bayesian data analysis, mixed modelling, cognitive decline
Using deep learning to improve genomic risk prediction of Alzheimer’s Disease
It’s well known that an individual’s genetic profile affects their risk of Alzheimer’s Disease and their rate of cognitive decline. This project aims to understand how deep learning may be applied to a wealth of existing genetics-related data in order to improve prediction of an individual’s risk of Alzheimer’s Disease. The focus will be on models and architectures that allow for the integration of other modalities (e.g imaging, cognitive tests). We will also explore how the resulting prediction relate to other established Alzheimer’s biomarkers.
Key areas involved in the research: machine learning, genomic risk prediction, cognitive decline
Modelling of Alzheimer’s Disease induced cognitive decline with Machine Learning
The primary goal of this project is to utilise machine learning and artificial intelligence (AI) techniques in order to develop novel methods for understanding the nature and magnitude of changes in cognition and behaviour associated with those variations in brain structure and function that are hypothesised to be causal in Alzheimer’s Disease. The focus will be on modelling the cognitive and behavioural trajectories preceding the diagnosis of dementia, as well as determining which measurable factors (demographic, genetic etc.) may influence such trajectories. Two important additional challenges will be to find the best way of dealing with missing data as well as to study the extent to which biases from repeated application of tests can be corrected mathematically. Finally, the project will seek to estimate potential benefits to brain-behaviour models associated with high-frequency cognitive assessments.
As a result, quantitative hypotheses will be generated that can subsequently be validated in independent data sets or with specific experimental programs, predominantly in learning and memory, involving functional neuroimaging as well as further psychopharmacological and cognitive challenges.
Key areas involved in the research: machine learning, Bayesian data analysis, modelling of cognitive decline
Context-specific Clinical Decision Support Leveraging Clinical Practice Guidelines
This PhD project will focus on identification of relevant clinical guidelines for a given clinical situation. In particular, the emphasis of the project will be to explore methods from information retrieval and natural language processing to identify key clinical practice guidelines to guide clinical decision making. It will require some modelling of the typical clinical decision-making process and understanding of key decision points in clinical care, in order to align clinical situations to particular steps or stages in a clinical guideline. It will explore the required depth of analysis of guidelines relative to the clinical situation – is it sufficient to retrieve a complete guideline, or do specific paragraphs or sentences need to be identified? How important is presentation of the background evidence in the published literature to support the clinical decision?
If suitable clinical data sets can be identified, the project will also consider how to use clinical data from medical consultations to trigger identification of a relevant clinical guideline at a relevant moment in a patient interaction, to support clinical decision making.
This project is based at RMIT University, working directly with Professor Karin Verspoor and Professor Lawrence Cavedon and also benefits from the context of the Natural Language Processing group at the University of Melbourne, including Professor Tim Baldwin and Professor Trevor Cohn, and their existing students and research fellows. Professor Baldwin is the Director of the training centre, while Professor Verspoor is the Deputy Director.
Key areas involved in the research: natural language processing, information retrieval, health informatics
Multi-document summarisation supporting clinical evidence review
This project will focus on summarisation of the background literature relevant to a given clinical question. Under the assumption that a relevant body of literature for a clinical question can be identified, the objective will be abstractive summarisation of that literature with sensitivity to the clinical question that is being explored.
The student project work will focus on natural language generation for the summaries, and consider adaptation of the summaries for different audiences based on readability, background knowledge, desired summary length, etc. to cater to different requirements.
This project also benefits from the context of the Natural Language Processing group at the University of Melbourne, including Professor Tim Baldwin and Professor Trevor Cohn and their existing students and research fellows, and Professor Karin Verspoor from RMIT University.
Key areas involved in the research: natural language processing, health informatics
Seizure Forecasting using Neural Networks
This project aims to improve performance of epileptic seizure forecasting. The outcome will be a system that provides estimates of the likelihood that a person will or will not have a seizure during a relevant forecasting time horizon. Artificial neural networks have demonstrated superior performance in seizure forecasting. This project will build upon our earlier work to develop new approaches based on convolutional neural networks, recurrent neural networks, and other approaches. There is potential to incorporate other sources of relevant data, such as time of day and movement tracking, as well as EEG.
The primary supervisor for this project will be A/Prof Lin Liu at the University of South Australia, but our PhD student will spend a minimum of one year physically based in Melbourne.
Key areas involved in the research: signal processing, machine learning, neural networks
Deep Learning for Fitting Neural Mass Models to EEG
This project aims to improve performance of seizure prediction by linking neural mass models to deep learning techniques. Neural mass models have been able to model and reproduce output signals that replicate different types of intracranial EEG recordings. Data assimilation with Kalman filters has been successful in investigating how variations in model parameters can lead to the different behaviours. However, to use these models for seizure prediction, high fidelity models are needed that incorporate mechanisms of parameter variation that are intrinsic to the model. This project will develop augmented neural mass models and deep learning and other machine learning techniques to fit the models to patient-specific iEEG recordings.
Key areas involved in the research: advanced signal processing, machine learning, neural modelling
Brain-Computer Interfaces for low-powered mobile/wearable devices
This project aims to develop a brain-computer interface system with emphasis on low power consumption and small memory requirements, without compromising on accuracy. The outcome of this project will be novel CNN models that incorporate spatiotemporal information, produce fast inference, and achieve small memory foot-prints so that they can be deployed in wearable devices or mobile computing platforms. The greatest challenge will be to maintain the required accuracy and performance while reducing the model size and the overall latency of the system.
Key areas involved in the research: signal processing, machine learning, neural networks, embedded systems
Multimodal Brain-Computer Interfaces for augmented reality and virtual reality applications
This project aims to develop a brain-computer interface system that allows people with speech, vision or motor disorders to interact in an augmented/virtual environment (e.g., selecting/manipulating an object of interest, walking/exploring a remote site). This AR/VR-based BCI will allow users to learn and adapt to the BCI in a more swift manner with countless possibilities of application scenarios, compared to the conventional case of a task-specific BCI (e.g. to control a robotic arm). Furthermore, AR/VR-based BCI systems will be safer to operate, can be deployed anywhere with minimal hardware constraints, and will provide an extremely efficient training platform to decode the brain’s intent. Brain-computer interfaces rely upon good recordings of neural signals, typically with scalp or intracranial EEG systems. This project will build a multimodal BCI system that augments the EEG signals through speech recognition and gaze tracking as a multimodal input to the system to further enhance the performance/accuracy of the system for the desired application scenarios.
Key areas involved in the research: signal processing, machine learning, neural networks
Information Extraction for Precision Patient Matching
This project focuses on information extraction over text sources including patient descriptions, literature, and clinical trial descriptions, to accurately map patient descriptions to relevant literature and clinical trials. Information extraction will take the form of both explicit span-based chunking and labelling (e.g. based on PICO criteria), and implicit generation of keywords and latent representations that capture key aspects of the corresponding document. Also part of this project will be the assessment of semantic compatibility between (labelled) text fields across different document types.