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 Industry Partners Seer Medical and Synchron along with 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 an Industry Partner and will spend at least 12 months of their candidature working on site at the Industry Partner’s facilities.

Our PhD students are working on the following projects:

Stream 1

Non-EEG approaches to Seizure Detection and Prediction

Collecting EEG data from patients for considerable periods of time is burdensome and highly inconvenient. Smart watches such as the FitBit (Google) device can provide multiple physiological data, including heart rate, temperature, and blood oxygenation levels.

Stream 2

Alternative single-switch decoder

This project aims to develop an alternative pipeline for implementing a single-switch brain-computer interface to the one currently used by Synchron by adopting state-of-the-art algorithms for motor imagery decoding. The availability of multiple decoding pipelines would allow participants to choose the approach with which they achieve the best performance. The decoding pipeline will be tested on data from multiple participants currently taking part in the trial.

An SSVEP-Based Multi-Modal Brain-Computer Interface with Augmented Reality

The aim of this project is to use SSEVP delivered via augmented reality to enable the user to interact with their environment.

Handwriting decoding

This project explores the feasibility of decoding from Stentrode data which character a participant is attempting to write. The project will primarily focus on understanding what features encode information relevant to handwriting and how this information is encoded across time. Those characters which can be decoded most robustly will serve as a basis for a real-time multi-class decoder.

Shared Control of a Robotic Arm Driven by a Noninvasive-Brain-Computer Interface

The limitations of brain-computer interfaces (BCIs) can be overcome by combining BCI outputs with intelligent automation. This project aims to achieve this by creating a noninvasive BCI for semi-autonomous control of a robotic arm and an endovascular brain-to-text BCI that utilizes language models. To evaluate this concept, the project employs ‘shared control,’ a semi-autonomous framework for BCI-controlled robots that combines user inputs with autonomous control signals. The advantages of this approach were assessed by having participants use a shared control system to reach objects with a robotic arm.

Two-Dimensional Continuous Cursor Control using a Brain-Computer Interface

This project, in collaboration with Synchron, aims to develop a brain-controlled computer cursor control system that will enable severely paralysed people to control smart devices using brain signals. The neural activity is captured by a minimally invasive endovascular Stentrode device that records the brain’s activity from within a blood vessel.

Stream 3

Improving genomic prediction of AD using AI

This project focuses on how Bayesian and deep learning methodologies can be applied to large-scale genetics datasets to predict an individual’s genetic risk of AD, focusing on architectures that enable integration of other modalities (e.g. imaging, cognitive tests).

Modelling of Alzheimer’s induced cognitive decline

This PhD project aims 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 hypothesized to be causal in Alzheimer’s disease.

Using modern Machine Learning for probabilistic modelling of dementia

This PhD project focuses on the probabilistic modelling of Alzheimer’s Disease (AD). As the most common form of Dementia, AD has inspired many works at the intersection of neuroscience, statistics, and computer science to develop methodologies and models able to help in the clinical diagnosis of the disease, or estimating the evolution of patients’ neurodegeneration and cognitive decline when viewing AD as a continuum. Multiple aspects characterize the disease and pose significant challenges in its modelling: these includes heterogeneity in clinical presentations, incomplete observation of the disease through potentially noisy biomarkers, uncertainty about the biological or genetic mechanisms involved in the disease, and in most cases, an unknown time of onset. Whether for estimating disease progression or establishing predictions of diagnoses, viewing the problem through the lens of machine learning and probabilistic modelling allows for more accurate depictions of the underlying nature of the disease and how it translates into observable biomarkers with noise inherent to real-world biological data and imperfect measurement means.

Stream 4

Literature-based Discovery (LBD) in Alzeheimer’s Disease

This project is focused on Literature-based Discovery, as applied in the context of Alzheimer’s Disease (AD). A paper has been submitted (under review) which examines large-scale predictive forecasting of novel links/relationships between genotypes and phenotypes in Alzheimer’s. This analysis is based on transformation of a large library of AD publications into a knowledge graph, and the application of link prediction methods to propose new links.

Multi-document Summarization Supporting Clinical Evidence Review

Summarising (often contradictory) results of multiple clinical trials into conclusions which can be safely implemented by medical professionals in their daily practice is a very important, but highly challenging, task. In this thesis, we first examine what constitutes a well-formed answer to a clinical question, and define modality (certainty) of evidence both from the linguistic perspective and that of biomedical sciences. Next, we present our framework for human evaluation of the reliability of generated summaries which is based on the aspects we outlined (biomedical entities, direction of effect, and claim certainty), use it to highlight issues with the current models, and examine the possibility of automating the summary evaluation using large generative language models. Following that, we present our multi-document summarisarion dataset which has several levels of inputs and targets granularity as well as rich annotation for the clinical evidence aspects we defined, and use it in several scenarios to test capabilities of existing models. Finally, we turn to the question of synthesing the input studies into conlusions, in particular, reflecting the direction and certainty of findings in summarie.

Negation in neutral language processing

This project focuses on the analysis of negation in large language models (LLMs), which has particular relevance in the context of medical evidence – effective analysis of inclusion and exclusion criteria and meaningful assessment of presence/absence of symptoms and diagnoses hinges on correct treatment of negation in clinical texts and biomedical journal publications.

Numerical Information Processing from Text

This project is tackling another identified problem in current NLP/LLM tools directly relevant for biomedical evidence analysis, which is numerical information processing. Rahmad has been actively characterising the requirements for numerical information processing particularly relevant to analysis of clinical trials, as well as characterising the limitations of LLMs in handling numerical information. He further participated in SemEval 2023 Task 7, focused on multi-evidence natural language inference for clinical trial data (NLI4CT). This has provided good insight into the interaction between inference/reasoning and numerical information processing, and will support Rahmad to contribute to improving the use of NLP for evidence synthesis.