2021 Overview

The partnership between ACEMS and the Sax Institute maintained momentum with multiple collaborations and engagements, and laid the foundation for ongoing engagements and collaborations in 2022 with ACEMS members.

Highlights for 2021 include:

• Two PIs from the Sax Institute, Kerrin Bleicher and Tina Navin Cristina, actively engaged with ACEMS members throughout the year.
• Four new research collaboration ideas were submitted by both the Sax Institute and ACEMS members in response to the Ideas Challenge.
• ACEMS and Sax Institute commenced collaboration on a new project, “Free-Text Field Data from Hard Copy and Online ‘45 and Up’ Study Responses”.
• Sax Institute team members engaged with ACEMS in relation to its Synthetic Data research project, consortium, and workshop.
• Sax Institute team members participated in the ACEMS Final Annual Retreat 2021 and gave a presentation on Industry Day.

ACEMS Ideas Challenge: Supporting the Sax Institute & Health

ACEMS Stakeholder Engagement set a challenge for ACEMS partners and members to propose ideas for collaborations, aimed to help partners by leveraging ACEMS capabilities. The Sax Institute and ACEMS responded with four different collaboration ideas, three being from Sax and one from an ACEMS PhD student.

Research Collaboration - Ideas Challenge Ideas

The Sax Institute submitted project ideas including:

• Free-text field data from hard copy and online ‘45 and Up’ questionnaires.
• Assessment of survey questions of time.
• Comparing utility of different dimension reduction methods (including manifold learning techniques) for high-dimensional health data, to capture important variability across many health/demographic measures.

Ultimately, the first project received funding to proceed, in a collaboration between a team at Sax and NSW Health (secondee) plus ACEMS Researchers and a new ACEMS PhD student.

Sax Institute and ACEMS Ideas Challenge Research Project “Free-text field data from hard copy and online ‘45 and Up’ questionnaires”

Free-text field data from hard copy and online ‘45 and Up’ questionnaires

Free-text collected from ‘45 and Up’ Study questionnaires is known to contain potentially identifying information, such as names and places. There are three components to this project. First, an analysis of the extent and type of personal/identifying information in the free-text across different questions will be conducted. Second, existing tools for the removal of free-text will be identified and evaluated. Third, a selected tool will be applied to ‘45 and Up’ Study free-text fields, and the outcome validated.

Synthetic Data (Research, Workshop, Consortium)

ACEMS Research Project: “Flexible generative models for creation of synthetic data”

Data privacy is key at the Sax Institute. Accordingly, the potential offered by advances in synthetic data creation, enabling safe access to data, is of great interest to Sax (amongst others).

To help Sax and address limits to current synthetic data methods/approaches, ACEMS has supported:

• Research into synthetic data - including a project funded by the Centre’s Research Sprint Scheme on “Flexible generative models for creation of synthetic data”.
   
  Read more

  ACEMS-funded research project - “Flexible generative models for creation of synthetic data”

  ×

  Motivation:

  • Many organisations recognise the benefits of sharing their data or using it more widely, but face limitations in doing so because of privacy and other restrictions. In these cases, an attractive alternative is the development of synthetic data, that is, datasets that have the characteristics of the target dataset but aren’t a real dataset.

  State of the Art:

  • Popular approaches to generate synthetic data include autoregressive models, generative adversarial networks (GANs) and variational autoencoders.
  • GANs generate synthetic samples by mapping random input to a complex probability distribution via a trained neural network. The training of a GAN follows an adversarial learning approach where the objective function of a second neural network is to accurately discriminate between the observed data and samples from the generator. Conditional GANs (CGAN) generate samples from a joint probability distribution given a discrete variable taking a particular class. SpaceGAN is an application involving spatial data, which uses a CGAN to generate samples for a geospatial location conditional on the observations of neighbouring geospatial locations.
  • An important aspect of generating synthetic health-related data is being able to evaluate the performance of the model. Since the GAN is implicit, in that it does not model the likelihood function directly or allow for finding latent variables of particular samples, its performance can only be evaluated by comparing the synthetic data to the observed data. Proposed measures for the difference in distribution between the two datasets include the Inception Score (IS) and the Frechet Inception Distance (FID). Two-sample tests based on hold-out test data have also been proposed.

  Limits to current approaches and opportunity for ACEMS to contribute with statistics

  • These approaches have recognised limitations and there is great opportunity to develop a more comprehensive Bayesian framework for evaluating and validating these models.
  • The above approaches have only been developed in the past few years and are still relatively new in practice.
  • Moreover, they are not well known in the statistics community. There is great interest in utilising statistical thinking to expand the theoretical and methodological framework for these approaches, to create statistical computational code to implement them, and to apply these approaches to the generation of datasets that are restricted by privacy constraints.

  Interest of ACEMS Industry Partners and Affiliates:

  • These issues are being faced, for example by ACEMS Industry Affiliate Member the Cancer Council Queensland in the development of the next phase of the Australian Cancer Atlas: in light of the time and permissions involved in aggregating the cancer data from each State Registry and associated linked covariate information from the ABS, the availability of synthetic data would facilitate faster research development. The development of tools to create synthetic data, particularly with complex features such as spatial and spatio-temporal structure, would be of great interest to these and many other government organisations, companies and universities (for example, for use in teaching). We are unaware of any university groups in Australia with open software, practical experience and research capacity to develop synthetic data for general purpose, particularly for structured spatial or spatio-temporal data, with associated measures of goodness of fit and privacy guarantees.

  Aims of this research project:

  • The aim of this research project is to (i) review existing literature for deep generative models for generation of synthetic data; (ii) develop the methods, software products and capability, and (iii) create a statistical framework for evaluation and validation of the generated data.
  • This will have immense value both for the ACEMS collaborators and for future research projects. It is anticipated that this work will lead to (i) a journal article in a high impact applied statistics journal, (ii) publicly available code and (iii) further collaborative research and development projects with government partners.
  • A longer-term aim is an ARC Linkage Project application with industry partners, including the ABS and the Cancer Council Queensland.

  Close
   
  
• Co-hosted knowledge-transfer events including a synthetic data workshop with guest speakers from multiple ACEMS nodes and other institutions.
• Formation of a synthetic data consortium, led by ACEMS, with external stakeholders including the Sax Institute (amongst several ACEMS partners, affiliates and others) to be involved and benefit from specific synthetic data use cases.

Synthetic Data Workshop

The Sax Institute and other government, industry and academic guests attended the ACEMS co-hosted Synthetic Data workshop with the Australian Data Science Network (ADSN) on 10 September 2021. This workshop featured presenters from and beyond ACEMS, as featured in the table popup above, and enabled knowledge exchange.

Other industry guests included Cancer Council Queensland, ABS, Services Australia, Integrity Systems Company / Meat and Livestock Australia, as well as additional guests who requested access to the workshop recordings.

To learn more, watch these workshop presentations here or click on the individual videos.

Presentations from Synthetic Data Workshop (now online)

Overview of Synthetic Data Generation by ACEMS PhD Student Conor Hassan

Watch Presentation

GRATIS: GeneRAting TIme Series with diverse and controllable characteristics by ACEMS CI Professor Rob Hyndman

Watch Presentation

Deep Learning Techniques for Dealing with Lack of Data by Associate Professor Richi Nayak, QUT

Watch Presentation

Synthetic data generation using moment-based density estimation by Dr Bradley Wakefield, University of Wollongong (NIASRA)

Watch Presentation

Generating artificial video data to train machine learning algorithms by Dr Anthony Paproki

Watch Presentation

Learning with Limited Spatio-Temporal Data: Generative Adversarial Networks (and Alternative Approaches) by Professor Fiona Salim, RMIT

Watch Presentation

Kirsten Jackson from the Sax Institute also delivered a presentation. As the Senior Program Development Manager for the SURE Program in the Research Assets Division, Kirsten works with SURE’s 25+ data custodians and 500+ data users (including researchers), to understand data needs and ensure the right people access the right data, and that data protection meets the highest data governance and security requirements.

Three other Sax Institute team members also contributed to workshop discussions, including PI Tina Navin Cristina, Richard Summerhayes (on secondment from NSW Ministry of Health), and Kerrin Bleicher.

ACEMS work on synthetic data is of interest to the Sax Institute, including for its SURE program, and further opportunities for ongoing collaborations will be explored in 2022.

What is SURE?

Sax Institute’s Secured Unified Research Environment (SURE) is an online workspace for analysing and sharing health and human services data from many sources, including hospitals, Medicare, cancer registries and Centrelink. SURE offers custodians a flexible way to share health data with the research community, while also giving researchers access to pre-approved linked data they can trust. This national asset is used by over 500 researchers and 25 government and health data custodians across Australia, powering important health research.

Learn more about SURE at the Sax Institute website or watch this video about “What is SURE?"

ACEMS Final Annual Retreat

Notwithstanding pandemic restrictions, ACEMS was pleased to welcome all partners including Sax (virtually) to the ACEMS Final Annual Retreat 2021 on Industry Day. Sax Institute’s Tina Navin Cristina delivered a presentation with one of her ACEMS Ideas Challenge Project collaborators, Richi Nayak from QUT.

ACEMS Ideas Challenge Project Will Enable Use of Valuable “Coded”/“Free Text” Data from the Sax Institute’s 45 and Up Study Research Asset

Partner Investigators

ACEMS thanks the Sax Institute and their team for their valuable collaboration and engagement in 2021, particularly enabled by our partner investigators:

• Tina Navin Cristina, Sax Institute’s Data and Analytics Manager, who returned from maternity leave in the second half of 2021 and was then engaged in the Ideas Challenge and subsequent project.
• Kerrin Bleicher, the Manager of Sax Institute’s 45 and Up Study, in its Research Assets Division. From the second half of 2020, Kerrin stepped in as acting PI when Tina was on leave, and has remained engaged since, including for a strategic research planning and briefing session in late 2021.

Tina Navin Cristina

Data and Analytics Manager

Kerrin Bleicher

Manager, 45 and Up Study, Research Assets Division