IntelliHQ Research Leads Posters Research Ideas Group outcomes Summary

AI in Healthcare and IntelliHQ

Opening - Mr Ron Calvert, Chief Executive of GCHHS

Mr Calvert warmly welcomed all attendees to the research symposium. He was both pleased and excited to see such a large and varied audience, all focused on improving patient outcomes through the application of AI to Healthcare. Mr Calvert recognised the growing importance of the field in modern times, coupled with the increasing ability brought by better hardware, software, and the growing talent base in the field.

He welcomed the participants from the 7 universities represented, across a variety of disciplines, and both recognised and emphasised the importance of a multi-disciplinary approach to improving healthcare. This includes the multiple specialty areas represented by the universities, the wide range of clinicians within the HHS, and the help of industry, from startups to multi-nationals.

Mr Calvert was very excited by the initiative of a Health AI workshop, and welcomed it as an important direction for both the HHS and Health and Knowledge precinct. It is an opportunity for the Gold Coast to achieve world recognition in this emerging and exciting field.

IntelliHQ overview - Dr Kelvin Ross, PhD

Dr Kelvin Ross. Co-founder, IntelliHQ.

IntelliHQ is a partnership between Gold Coast Hospital and Health Service, Universities and Industry to transform healthcare through the practical application of next generation technologies. Central will be artificial Intelligence, which is already creating a 4th industrial revolution globally, and will markedly and continuously improve both patient and health system outcomes in the near future.

IntelliHQ is bridging the innovation gap for the Heath service and Industry, providing a pathway for cutting edge innovations to be rapidly introduced and trialed in the health system with maximum benefit and minimum risk. This will occur by bringing together the 4 pillars required, namely industry (startups, SMEs and multinational partners), research (across multiple institutions and specialties), education (including a training academy for both AI experts, and executives working with data scientists) , and Health data (via both clinical research, and the data platform in the next presentation).

An experienced and diverse advisory board is already well advanced in turning this vision in to a reality. 

Data Platform - Dr Brent Richards, FRACP FCICM

Central to the IntelliHQ direction is a data platform. Intensive Care has considerable data already available; at Gold Coast and across Queensland through the iMDSOft Metavision programme. With this there are over 25,000 patient records across Queensland with 1 minute resolution data, providing already a rich treasure trove of data once combined. Additionally, GCUH ICU is working to get a data stream from the GE monitor system gateway – a device at the centre of the monitor network, capable of giving data points every second from the monitors, along with waveforms at 60Hz and ECG at 240Hz. As proven already by current research on Heart Rate variability, this data can and will provide a rich stream of clinical insights, able to be tapped for both retrospective and eventually real time analytics.

There are many similarities between a self driving car journey and an ICU patient journey – knowledge about both the patient and disease prior to intervention, the nature of the intervention (e.g. operation), and then a fully monitored journey through intensive care, getting a patient quickly and safely through this critical period on the journey to full recovery.

Therefore many of the current AI technologies (hardware, software) are applicable to the ICU patient journey. With considerable data already available and more to soon follow the time is now to work to improve the safety and quality of this journey.

Additionally, what we learn from the ICU will also be useful in the hospital wards and later in homes, particularly as the technology allows streaming physiology data from these other venues.

 AI has the opportunity to markedly improve Healthcare. Building a data platform based in ICU is a critical stepping stone in this journey.

Lead research group presentations

University of Queensland School of Information Technology and Electrical Engineering

Author_ Prof. Xue Li

We are an active research group in School of Information Technology and Electrical Engineering, the University of Queensland. We have one full professor, three post doctors, eleven PhD candidates. Our research focus is machine learning on medical data analysis including precision medicine, medical knowledge graph, side-effect analysis for drug interaction, computations of patient similarity, severity prediction, medical image processing, etc. Our ultimate goal is to create automatic monitoring and support systems to help medical doctors provide better services to ICU and clinical patients. By using our system, experiences of doctors are enriched, the accuracy of doctors’ decisions is enhanced, the utility efficiency of facilities is improved.  We have published a series of high-quality publications in top journals and conferences including IEEE TKDE, ACM TKDD, Decision Support Systems, ICDM, ECML, etc. We are looking for collaborations with research partners.

Griffith University Institute of Intelligent and Integrated Systems

Lorem ipsum dolor sit amet, consectetuer adipiscing elit, sed diam nonummy nibh euismod tincidunt ut laoreet dolore magna aliquam erat volutpat.

Griffith University School of Information and Communication Technology

Professor Bela Stantic, Head of School
Director “Big Data and Smart Analytics” Lab – IIISSchool of Information and Communication Technology

Big Data and Smart Analytics lab, within the Institute for Integrated and Intelligent Systems, led by Director Professor Bela Stantic consist of 6 young experts fully committed to lab projects with expertise in the areas of data management, data mining, machine learning, and pattern recognition. Additionally, six research fellows are working on different funded projects. There are also several senior researchers being involved with fraction of their time as well as more than twenty PhD students and several visiting Scholars and Fellows from Big Data lab from China.

Funding for projects are from various sources, including local governments and industries. We are also targeting competitive grant applications for ARC and NHMRC from the federal government.

We aim top international conferences and journals, including SIMOD, ICDE, ICDM, AAAI, IJCAI, KDD, ECMLPKDD, VLDBJ, TKDE, TKDD, etc.

We have our dedicated cluster to support parallel computing and in-memory computing dealing with challenges in Big Data applications, such as Social Media Analytics and Big Medical Data Analytics.

Bond University Business School

Bond Business School, Bond University and intelliHQ

Bond Business School Leads: Dr. Bruce Vanstone and Dr. Adrian Gepp

The Bond Business School is continuing to increase its capabilities in Big Data Analytics, both in teaching and research. Healthcare is an important application domain for these skills and we support the IntelliHQ initiative with our Executive Dean sitting on its advisory board. We are interested in the application of big data techniques in healthcare and statistically rigorous evaluation of the resulting models. With expertise in financial modelling, we are very comfortable with handling multiple high-frequency time-series data streams – something that intelliHQ now has available for medical data.
We have already established a research partnership with intelliHQ. In additional to preliminary work on automating the patient referral process, we are completing a project this year about improving the currently used ICU severity scoring system by incorporating high-frequency data (see separate abstract). This research is being conducted by an Actuarial Honours student. In the future, we will have a growing number of such students with high-level quantitative and statistical skills looking for applied research projects.

Gold Coast University Hospital Intensive Care

The Intensive Care Unit at Gold Coast University Hospital (and previously Gold Coast Hospital) has over 20 years of research experience. This includes local investigator led studies, multi-centre studies (both locally and externally led), larger ANZICS CTG (Clinical trials group) studies and point prevalence studies, and a range of commercially led international pharmaceutical and device company trials.

The studies have included a very broad range of ICU topics, particularly sepsis and antibiotics, as well as dialysis, ventilation, pneumonia, nutrition, and DVT prophylaxis. Current major locally led research is based around coagulation in trauma, heart rate variability, and outcome prediction.

The ICU research team includes 4 research nurses and associates, assisting and assisted by 10 consultants and 150 nursing staff. WE have a Clinical Information System, Metavision, with 3 years of minutely data from GCUH, and a further 3 years of data from GCH. We will shortly be expanding this data to every second.

We are looking to expand in all areas of ICU prediction, including prevention of complications, as well as developing more precise medicine for individual patients.

Queensland University of Technology, Curtin University, Southern Cross University, eHealth Queensland

At the workshop a number of other institutions and research groups were represented.

Dr Ray Johnson from QUT discussed both their interest and availability im the area, with strong departments both in data science and mathematics, as well as excellent 3D printing and Robotics groups. 

Professor Luke Haseler from Curtin University welcomed the opportunity for collaboration, noting a strong exercise science programme at Curtin, a new medical school, and a partnership with CISCO. He celebrated the successes of the heart rate variability group to date, saying it was a model for successful partnerships between University and hospital.

Prof. Dian Tjondronegro from SCU introduced his group. They all have substatntial experience and interest in the health and wellness measurement space, particularly home care, and welcome ongoing collaborations.

Professor Peter Bevan from eHealth Queensland and QUT welcomed the breadth of interest and engagement at the workshop. He noted that the pathway for startup engagement and acceleration as in place through IntelliHQ adn GCHHS is the model eHealth QLD are using, and encouraged all to join in on the pathway.


Temporal Similarity Computing on ICU Data

Title: Temporal Similarity Computing on ICU Data

First Author: Suresh Pokharel

Affiliation: ITEE, UQ 

Abstract: Finding the similarity among the ICU patients creates many benefits such as case based patient retrieval, recommend similar treatment, unearth similar patients group. However, similarity computing in ICU data is not an easy task. ICU data are complex and has the following properties:  heterogeneous, temporal, sparse, irregular and multivariate. Computing such complex data brings many challenges; first, representation of such complex data into machine processing form without losing its information; second, how to make them comparable; last but not the least, selecting features for similarity discovery. To address these challenges, we propose Resource Description Framework (RDF) technology for representing ICU data. Then, we will enrich data by adding temporal relations, by adding abstract information, etc.  Finally, we will analyze trends, values, abnormality information, node and link, etc. for finding the similarity between patients.

Big Data and ICU scoring systems

By James Todd, Brent Richards, Bruce Vanstone and Adrian Gepp

Bond Business School in conjunction with GCUH

Severity scoring systems are used in intensive care units for stratifying patients in clinical research and benchmarking ICU performance. A variant of the Acute Physiology and Chronic Health Evaluation (APACHE) system is used in Australia for benchmarking purposes – the APACHE III-j. This and other major scoring systems have been developed under an old paradigm of minimum data collection, while the current paradigm is to use all useful data. The APACHE III-j system uses the worst observations in the first 24-hours of a patient’s ICU stay, ignoring the rest of the distributional information. We hypothesise that scoring system performance can be improved by adding variables that capture this ignored distributional information. To test this hypothesis, the APACHE III-j system will be replicated and compared to a modified version that adds metrics describing the distribution of an underlying physiology variable utilising high frequency data.

Towards a Workflow-Centric, Context-Aware Clinical Application Framework

Title: Towards a Workflow-Centric, Context-Aware Clinical Application Framework
Authors: Dr. Zoran Milosevic (Deontik, Australia); Dr. Davide Sottara (Mayo Clinic, USA; Arizona State University, USA); Dr. Matt M. Burton (Mayo Clinic, USA; Arizona State University, USA) ; Prof. Robert A. Greenes (Arizona State University, USA)

We outline an approach for implementing application support for clinical workflow that is aware of dynamic context of healthcare delivery. This context captures patient specific information, previous and future activities of clinicians involved in patient healthcare and supports clinicians’ cognitive models. This application framework leverages some of the concepts and models that have been used to formalize a major, multidisciplinary clinical workflow redesign project at a major US hospital. We present our results and future work, and in particular options for modelling context of healthcare delivery using a combination of a formal ontology to capture existing semantic agreements on clinical workflow and artificial intelligence as a way of dynamically inferring context features.

Medical Knowledge Graph for Genomics in Critical Care

Title: Medical Knowledge Graph for Genomics in Critical Care

First Author: Mingyang Zhong

Affiliation: ITEE, UQ 

Abstract: Genetic factors drive key aspects of inflammatory outcome, and identifying the genes that drive inflammation is meaningful in critical care. In medical and genomics areas, researchers usually apply meta-analysis based approaches that are the statistical procedures for combining data from multiple studies to identify common effects. Data mining and machine learning based solutions in data science for applying genomics to clinical practice and providing practical strategies to help the health system planners and policy makers can be naturally modeled using Knowledge Graph. In this research, we investigate multi-sourced data and develop a reasoning model that detects potential association among genes, diseases and symptoms and a ranking model that finds objects (patients, diseases, etc.) with similar attributes and provides recommendations based on evidence to decision makers. 

Dynamic Illness Severity Scores Prediction in ICU Using Multi-task Long Short-term Memory Learning

Title: Dynamic Illness Severity Scores Prediction in ICU Using Multi-task Long Short-term Memory Learning

First Author: Weitong (Tony) Chen

Affiliation: ITEE, UQ 

Abstract: Considering trajectory of an Intensive Care Unit (ICU) patient as a dynamical system, Multi-task Long Short-term Memory Learning (MTLSTML) method is investigated to learn the course of patient encounters in the ICU. Data extracted from MIMIC-III [1] of 38,645 adults patients in EMR (Electronic Medical Records)  over more than ten years. The objective of MTLSTML model is to dynamically forecast ICU patients’ Illness Severity Scores at user-specified time intervals base on ICU expert-defined sequential measurement data which include vital information, lab test results, drugs used and interventions. Our experiment results demonstrate that the hourly-based dynamic Illness Severity Scores Prediction (APACHE II, SAPA, and SOFA) not only expeditiously forecast the potential outcome of critically ill ICU patients, but also comprehensively evaluate the delivery of care to patients. It offers significant improvements over traditional clinically-used scores.

ICU-Acquired Weakness Classification by Deep Learning on Dynamic Networks

Title: ICU-Acquired Weakness Classification by Deep Learning on Dynamic Networks

First Author: Xingjuan (Miranda) Li

Affiliation: ITEE, UQ 

Abstract: ICU-acquired weakness causes significant functional impairments on the survivors of critical illness, leading to the slow and incomplete recovery. ICU-acquired weakness can be classified into three categories: critical illness polyneuropathy (CIP), critical illness myopathy (CIM), and critical illness neuromyopathy (CINM). It is often difficult to distinguish this myopathy from critical illness polyneuropathy simply by means of the bedside examination because both conditions are manifested by limb and respiratory-muscle weakness and retained sensory function. However, ICU-acquired weakness classification is valuable, since weakness in survivors of critical illness is common and is associated with long-standing consequences that dramatically affect recovery.

Moreover, as survival rates among patients in the ICU increase, ICU-acquired weakness will have increasing relevance for care providers outside the ICU. In this study, we attempt to use deep learning and big data approach to improve and support the clinical diagnosis of ICU-acquired weakness. Specifically, we propose a Deep Learning on Dynamic Networks framework to classify ICU-acquired weakness.

Applying real-time analytics to data streams in digital health

Title: Applying real-time analytics to data streams in digital health: reducing costs, detecting data quality issues and improving patient care
Author: Dr. Zoran Milosevic (Deontik, Australia);

We describe the benefits of real-time analytics, specifically complex event processing technology, in addressing a number of challenges in digital health applications. We focus on three uses cases. The first use case is a real-time detection of unusual or inappropriate laboratory orders, the problem which leads to significant and unnecessary costs to many healthcare providers. The second use case is the detection of potential data quality issues associated with source systems in pathology labs, by using a novel idea of applying syndromic surveillance method to the legacy clinical data streams, achieved through statistical analysis of pathology messages to identify “outliers”. The third use case is about supporting clinicians in making timely decisions regarding patient care, taking into account a combination of real-time information about patient conditions and their existing medical conditions taken from electronic health records. We demonstrate how we used the EventSwarm software framework for complex event processing to support this real-time analytics and will discuss a number of future research directions.

Treatment Recommendation by Observational Study

Title: Treatment Recommendation by Observational Study

First Author: Xin Zhao

Affiliation: ITEE, UQ 

Abstract: A randomized control trial (RCT) is a type of scientific experiment which aims to reduce bias when testing a new treatment. Unfortunately, RCTs are often infeasible due to practical and ethical limits. An observational study is an alternative solution to infer treatment effect. An observational study draws inferences from a sample to a population where the independent variable is not under the control of the researcher. This is in contrast with RCTs. In this presentation, we proposed a new method to use observational data to mimic RCTs.  Our method can find a subpopulation, within this subpopulation two goals can be achieved: 1) Balanced distributions in the treatment and control groups – Fair comparison in RCT; 2) Maximized difference of outcomes between treatment and control groups – Well designed RCT. 

Causality discovery on drug-drug interaction

Title: Causality discovery on drug-drug interaction

First Author: Sitthichoke Subpaiboonkit

Affiliation: ITEE, UQ 

Abstract: An adverse side-effect of drug-drug interaction from concurrent drugs consumption is an acute problem because it can significantly cause morbidity and death. The high cost to solve this unexpected adverse event is inevitable. Drug-drug interaction can be prevented if it is discovered before market approval. Unfortunately, it is very difficult to practically perform because the possible combination of drugs usage needed to be analyzed is extremely large, and new drugs can be experimented to find interaction with existing drugs by performing in vitro and in vivo methods, complicated, costly, and time- consuming methods. Computational methods from observational drug-drug interaction usage data are introduced to scope the cases to be then confirmed the drug-drug interaction with in vitro and in vivo methods; however, they are mainly based on association analysis, which does not imply causation. In this research, we will introduce the practical computational causality discovery on drug-drug interaction enhancing by available domain knowledge in drug side effect.

Patterns of Heart Rate Variability in Traumatic Brain Injuries across the first 24hrs of Intensive Care Admission

Tegan Roberts1,2, Brent Richards1,2, Luke Haseler1,3, Matthew Wells1
1Griffith University, School of Medical Sciences,
2Gold Coast University Hospital, Southport,
3Curtin University, Western Australia

Traumatic Brain Injury (TBI) is one of the leading causes of death and disability in young adults and the elderly population1, and despite the improvement in resuscitation and critical care of these patients, predicting patient trajectories remains difficult to establish, particularly within the first 24 hrs. One of the most detrimental and immediate consequences following TBI is sympathetic nervous system hyperactivity, causing increased cerebral inflammation, hypertension and hypoxia.
The natural variation of the hearts beat-to-beat intervals, Heart Rate Variability (HRV) can be derived from sequential electrocardiogram (ECG) R-R measurements. An individuals HRV is naturally variable, with decreasing variability identified during periods of illness or disease2. HRV has demonstrated to reflect changes in the autonomic nervous system in a variety of clinical settings including subarachnoid haemorrhage, trauma and brain injuries3.

ICU patient presentations are highly variable, as are their responses to therapeutic interventions, thus clinicians are always looking for newer indicators that help further define severity of illness and likelihood of deterioration. The evident separation between survivable and non-survivable HRV parameters highlight HRV as a promising ‘e-biomarker’ of injury severity in TBI patients and may be useful in predicting patient trajectories and with end-of-life decisions.

Evaluation of RR interval Estimation Techniques

Meghan McConnell, Belinda Scherwin, Brent Richards

RR intervals are increasingly being used to understand different physiological states. For this to be useful the R wave peaks must first be accurately detected in an ECG. 

Therefore the current available algorithms were tested (Kubios, Pan-Tompkins, Hilbert transform, K-means) along with two new methods I wrote; a modified Hilbert transform, and an ensemble technique.

On testing both Kubios and standard Hilbert transform performed poorly. Best performance was a modified Hilbert, and an ensemble technique.

Further testing on a larger sample is needed. IF the results are confirmed then the Kubios method should be replaced by the ensemble technique.


Can We Trust AI In Healthcare?

Dr Kelvin Ross

Chairman, KJR / Adjunct Assoc Professor, Integrated and Intelligent Systems, Griffith University


Recent rapid advances in machine learning technologies hold great promise towards improving healthcare outcomes.  Applications such as image recognition applied to disease diagnosis, classification used in abnormality detection and early warning scores, and reinforcement learning optimising personalised drug dosing are a few recent examples of significant advances.  


The adoption of machine learning technologies though has several trust and assurance challenges that must be addressed to build confidence amongst healthcare providers to gain widespread clinical adoption.  In this presentation I discuss several key challenges:

  • Assessing prediction accuracy
  • Augmenting clinicians for optimal accuracy and reduced human errors
  • Understanding data errors, overfitting to specific protocols, and data set bias
  • Explainability of AI decisions to support clinical judgement
  • Patient data privacy
  • Incorporating AI into a quality product within clinical workflow
  • Regulation of AI technologies


Research Ideas

Using neural networks to recognise normal vs abnormal beats, and the presence of arrythmias

Monitoring patients post discharge – ward and home – looking for deterioration

A programme of Image analysis, including Xrays.

Using facial expression recognition to look for pain

Looking at ideal workflow for taking X-rays, particularly for emergencies – what to complete, what to delay. 

Further work on heart rate variability, expanding patient cohorts and diseases, including using it as a predictor of complications.

Fingerprint analysis for disease detection

Using AI to determine severity, predict mortality, predict complications, monitor patient progress

Precision medicine treatment recommendation/decision support system. Concentrate on explainablity.

Predict which is the best antibiotic to use for an individual. Particularly, identify those patients who will only require simple antibiotics (i.e. risk of resistant organisms is negligible); thus preserving powerful antibiotics. 

Look at multi-modal monitoring and the interactions between these signals, with a view to additional feature recognition from combined signals.

Investigate medical decision making based on prbability reasoning

Develop per patient clinical decision support for pathology ordering

Look at options around surgery in the obese patient

Determine optimum individual care pathways for the post-operative period.

Understand cognitive load in the ICU setting. Recognise maximum decision making capacity and level of required recommender systems.

Use camera technology to visually identify moulds.

Predict which individual patients are appropriate for fast track care, with shorter ventilation and in-ICU times.

Use data to more accurately propensity match patients from existing databases to better determine appropriate therapies in patient groups. This would include feature selection on both known predictors, and discovering new predictors.

Better determine individual drug dosing, based on known parameters of pharmacokinetics, and improving these predictions with a wider physiological and pathology data set.

Develop scores for early recognition of the deteriorating patient, both in ICU and in the wards. Use this to also decrease re-admissions.

Using reinforcement learning on data streams to better understand and predict patient journey.

Activity monitoring, including: sedation in ICU, speed of post-operative recovery, monitoring of imporvement post-discharge. 

Group outcomes


Leader – Tegan Roberts, GCUH/GU

Heart rate variability is becoming a promising e-biomarker for a variety of clinical conditions, including in Intensive Care. The work to date has been promising in Acquired Brain Injury, and there are plans to expand in to concussion.

Determining HRV requires IoT, signal detection, validation and measurement, along with analysis of these measures in multiple dimensions alongside clinical data – thus an ideal programme for AI in ICU and beyond,

First project will be to develop a NN recognition of ECG (similar to the Andrew Ng Stanford project) to better rhythm-based signal-to-noise in a recording.

The programme will then extend the use of the technology to all ICU patients, looking at ABI, sepsis, sedation and weaning, in addition to looking further at ABI by researching the concussion space. Additionally, as well as the current freqency-based RR analysis, further discovery will occur both in the entropy space, and application of Nueral network learning to the ECG traces.


Finalising. Working on both imaging workflows and image recognition.

Early warning of patient detrioration

Leader: Ping Zhang

Finalising plans to develop early warning scores for ICU, wards, and home settings.

Decision Support

Being finalised. The AI needs to converted to timely decision support.

Intelligent Dashboards

Finalising. User interface will be key to delivering AI to clinicians for timely and appropriate action.

Workshop Summary

With almost 60 people attending throughout the day, this workshop was the largest gathering of AI researchers in ICU every seen in Australia, and potentially in the first world.

From the opening by the GCHHS CEO, through great overviews on IntelliHQ and from the research leads, the day started with considerable energy and focus. From there some excellent posters were presented on research already in progress, and then a broad array of potential research areas were proposed – a list that was hard to narrow down.

Around the most popular topics 5 groups formed and worked hard over the course of the afternoon – initially to develop their first research proposal, and from there expand in to a research programme; all the while keeping in mind funding opportunities and commercialisation.

All groups created very well crafted programmes, elected group leaders, and agreed to keep meeting and working towards their combined goals. Additionally everyone agreed we should make this AI in ICU research meeting a regular event, starting 6 monthly.

Thank you everyone for an incredible day of energy and discovery. Our future as a combined research group is very strong, and we will become one of the leading AI research groups in ICU.