Skip to main content

Explainable and Robust Deep Causal Models for AI assisted Clinical Pathology

Primary supervisor

Lizhen Qu


Research area

Vision and Language

We are living in the era of the 4th industrial revolution through the use of cyber physical systems. Data Science has revolutionised the way we do things, including our practice in healthcare. Application of artificial intelligence/machine learning to the big data from genetics and omics is well recognized in healthcare, however, its application to the data reported everyday as part of the clinical laboratory testing environment for improvement of patient care is under-utilized. 

Millions of patient test results are generated each year by each medium-sized clinical laboratory, like Monash Health Pathology, in various formats including images, texts and numeric values. The study of these unstructured data in the pathology laboratory information system (LIS), together with the information from the electronic medical records (EMR), will offer us opportunities to provide better quality patient care at reduced cost. Data science will promote “4P” in medicine (predictive, pre-emptive, personalised and participatory), which includes earlier diagnosis, determine risk profiles, predict outcomes and involve patients. This is particularly important for specific groups where prospective studies are difficult to conduct, such as in infants and pregnant women. For example, gestational diabetes (GDM) is associated with increased pregnancy and birth complications and it affects up to 30% of pregnancies in Australia and 45% worldwide. Predictive models based on Artificial Intelligence (AI) techniques will enable risk stratification to implement personalized medicine for prevention and earlier diagnosis of GDM.

One of the challenges in using AI in knowledge discovery in Pathology is how to integrate health data from various sources and in different formats (image, text and numeric values). Up to now, AI, especially deep learning, techniques in pathology are used mainly to identify visual patterns in medical images, ignoring useful information in other modalities, such as clinical notes from EMRs and omic data from biobanks. Images alone are insufficient to study the causal effects of diseases. A further challenge lies in building robust models to discover novel knowledge from the integrated data. Lastly, prior models often appear as black boxes such that they lack explanations for why and how a particular prediction is made.

To address these challenges, this Ph.D. project aims to combine causal analysis with deep learning for knowledge discovery using all available data in various formats, in collaboration with Monash Health Pathology. It will start with building a knowledge system to integrate big data from various sources provided by Monash Health Pathology. Data fusion and association analysis will be conducted using deep learning techniques, known to work well for data representation and association learning. As deep learning is vulnerable to spurious correlations, novel causal discovery and inference methods will be developed to identify and reason over causal relationships among all associations from fused data. As the number of causal relationships is usually much smaller than that of associations, the proposed techniques will achieve explainability by making causes and effects interpretable to pathologists

We envisage that the combination of causality and deep learning will lead to the promotion of “4P” in medicine. The explainable models will enable pathologists and clinicians to formulate and verify their hypotheses more efficiently, predict risks more reliably, and identify previously unknown risk factors from a vast amount of data in a reasonable time, leading to better health care for patients.

The project aims to

  • Design and develop a system to integrate pathology data in different formats from various sources.
  • Develop cutting-edge deep causal models to identify causal relationships between risk factors and diseases by using data from Monash Health Pathology.
  • Devise the models that can produce explanations that are easily understandable by health professionals.

#digitalhealth #datascience #AI #ML #NLP #pathology #biochemistry #multimodal #analytics #healthcare #health #medicine #healthinformatics


Learn more about minimum entry requirements.