The industry problem/opportunity
World-leading cancer research, education and treatment institution Peter MacCallum Cancer Centre provides co-ordinated and comprehensive cancer care.
As with many large healthcare organisations, even those with a comprehensive Electronic Medical Record, patient data is stored across a number of systems. The separate mechanisms for capturing and storing patient data in a range of formats can limit the utility of these comprehensive datasets in research and development activities. Without the optimum infrastructure to manage data sharing, projects can take longer, and cost more.
The research
In 2022, the Digital Health CRC (DHCRC) brought together Swinburne University of Technology (Swinburne), Peter MacCallum Cancer Centre (Peter Mac) and Propel Health AI (Propel) for a collaborative data management project in two phases.
The first phase involved testing the Propel platform, which aims to provide a cloud-based analytics platform to accelerate clinical research and development workflows.
Aggregating multimodal datasets from selected source systems, the Propel platform ensures data is harmonised to an interoperable data model (FHIR), shared securely with controlled access to data contained within the platform’s Trusted Research Environment (TRE), and maintains privacy through the de-identification of all health records.
Professor Christopher Fluke from Swinburne University of Technology says this initiative presented a unique opportunity to bring a diverse data set together on one scalable platform for data-driven research and development.
“There is an ever-growing breadth of medical data but its form and disparity can make it difficult to use to drive tangible research outcomes. This project allowed us to consider the ethical and risk considerations of using a diverse data set and then assess how data and AI can help improve decision making outcomes for clinicians and patients.”
In response to rapid developments in artificial intelligence (AI), much research effort is being invested in developing potential solutions that could be applied to improve the efficiency and productivity of healthcare systems. In this vein, the second phase of the project aimed to develop an AI model to support clinical workflow in radiology reporting, helping clinicians to identify abnormalities in a time-efficient manner.
The solution and outcomes
Initially, the project was focused on demonstrating the capability of the platform. This included bringing select datasets together from different sources (e.g. electronic health record and imaging, lab, genomics and research systems) then linking the data, harmonising the data for interoperability, and aggregating de-identified data with the appropriate governance and regulatory oversight.
In the subsequent phase, focus shifted to validating the efficacy of the platform in the development of an AI model to support a clinical reporting workflow in radiology. The aim was to co-design a lightweight large vision and language model (LVLM) that aids radiologists in report writing and clinical communications as a proof-of-concept for improving existing practice and care.
The project team developed and trained the AI model on Swinburne’s Ngarrgu Tindebeek supercomputer using an open-source dataset as a cost-effective approach for rapid prototyping and experimentation. On the open-source data, the lightweight LVLM achieved performance outcomes approaching those of much larger state-of-the-art solutions. The proof-of-concept model was then deployed into the Propel platform for pilot testing on clinically-realistic data by select radiologists from Peter Mac.
As expected, preliminary evaluation of the prototype found that while the tool could analyse CT imaging, it could not identify specific abnormalities. This was due to the known mismatch between the open source data used to fast-track training and clinically-realistic data, such as the limited vocabulary overlap between the training set and the language used in radiological reporting. The validation step did show that the tool could achieve consistent inference speeds and provide useful outputs. These early outcomes confirm that further fine-tuning of the AI model on clinically-relevant, domain-specific datasets is required to achieve acceptable performance.
The impact
This research has demonstrated the Propel platform’s capability to securely capture and aggregate select patient data for a large healthcare organisation in Australia, allowing researchers to access consolidated, de-identified, interoperable data in a secure manner.
Its utility in enabling data-driven digital health research and development has been tested with the first proof-of-concept, leveraging the capabilities of the platform in the deployment, testing and validation of AI models for clinical and operational decision support. The results and learnings from the evaluation of this prototype’s effectiveness, accuracy and user experience can now be used to consider further development of radiology-specific foundation models addressing identified areas for improvement.
Peter MacCallum Cancer Centre Cancer Imaging Specialist Kwang Chin says, “The current speed and progress of AI in radiology requires all of us to change established paradigms. AI is now capable of processing vast imaging datasets with speed and consistency, allowing radiologists to prioritise complex diagnostic imaging and focus on more patient centred care.”
The evaluation of the prototype within the Propel platform highlights that access to real-world clinical data is critical for achieving reliable clinical accuracy in production-ready AI tools for digital health. This demonstrates the importance of secure data platforms with appropriate compute power for activating real-world data for model development.
Ultimately, solutions that unlock large health datasets to support collaborative digital health research under appropriate regulations will encourage new innovations in clinical practice, accelerating R&D, reducing operational costs, and improving patient outcomes.
The insight
Digital Health CRC Chief Executive Officer Annette Schmiede highlights the value of this collaborative project as an opportunity to transform how digital health data is stored and shared for the long-term growth of the health and medical research sector.
“Active use of every Australian’s health care data to support personalised healthcare, and use of linked longitudinal data to support population and public health research and development, will uplift the quality of life for every Australian,” she says.
“Given the unstructured and fragmented nature of health data and its exponentially growing size, AI is poised to play a key role for enabling and empowering health organisations to tap into the potential of that data to deliver real value for patients and the health system.
“Unlocking Australia’s healthcare data should be a national priority, for both the benefit of all Australians and to re-position Australia’s R&D capability and infrastructure as a sovereign asset.”
