AI for biology and health
Technological advances have made it possible to collect an enormous amount of clinical data, ranging from genes to medical records.
Our research aims to step away from the one-size-fits-all approach to healthcare, and instead develop methods that tailor disease diagnosis and treatment to specific individuals. To this end, we are running a number of projects that seek to:
- Understand the genetic and genomic basis of disease
- Identify biomarkers that can predict disease severity or treatment response
- Identify distinct sub-populations within diseased populations
- Develop algorithms that aid clinical decision-making
Our key focus is on:
Our Biomarker Discovery projects focus on using blood-based gene expression and gut microbiome signatures to establish non-invasive diagnostic tests for complex biological disorders. This work not only involves finding biomarkers to make health predictions, but also involves developing new methods for abstracting meaningful biomarkers from noisy health data.
AI for understanding
Our AI for Understanding projects focus on developing new workflows and new algorithms that make it easier for biologists and clinicians to interpret AI decision-making. This involves engineering abstract features that reflect domain-specific knowledge, as well as using deep learning to provide personalised risk scores.
Our Data Integration projects focus on bringing together vast amounts of clinical data into a single analysis. These projects include combing multiple views of the same patient, as well as joining data collected for different populations across different hospitals. To this end, we are working to identify the biological and technical factors that make these different data sources unique, and to develop new algorithms that make them comparable.
We have partnered with global experts in autism research to identify blood-based biomarkers from autism spectrum disorders. In this work, we seek to elucidate the molecular foundation for the spectrum, and to develop a reliable diagnostic test based on gene expression signatures.
If we intend to translate machine learning algorithms into clinical practice, we need a way to understand how algorithms make decisions. That’s why we’re developing new algorithms that provide a clear window into ‘black box’ AI. Unlike competing methods that can only tell the analyst if a patient has a disease, we have focused on creating new methods that can tell the analyst why a patient has a disease. By applying DeepTRIAGE to a number of health biomarker data sets, we’re gaining an understanding of the variability within diseases, and using this information to stratify patients into diagnostically relevant sub-groups.
In addition to developing algorithms that predict the presence of a diseased tissue, we’re also developing algorithms that can detect any departure from normal tissue function. These ’tissue detectors’ are able to identify any kind of unusual biological activity, flagging these patients for further evaluation by clinicians.
One major barrier in disease modelling is the vast amount of data available. This includes data collected from different assays, from different patients, and from different research centres. We’re developing new methods to bring these multiple data types together, allowing us to describe how different modes of clinical data influence one another.
Electronic medical records
In addition to health biomarker data, we’re using electronic medical records to develop algorithms that support clinical decision-making by ‘reading’ physician notes. We’re currently building models that predict future injury from past medical histories, as well as models that differentiate between ambiguous clinical presentations. As part of this work, we’ve partnered with medical practitioners within the community to ensure that our algorithms solve problems that are directly relevant to everyday clinical practice.