The technology uses deep learning to analyse chest scans, identifying patterns humans can’t see.
ResearchGate: What inspired this technology?
Lyle Palmer: The mathematical basis of our approach, deep learning, was inspired by attempts to model how the human brain learns. Although the fundamental ideas behind deep learning are not new, dramatic improvements in computing capacity and access to large datasets have enabled the widespread use of deep learning approaches in industry and academia over the last five years. Its use in medicine is currently at a very early stage.
Our specific use of deep learning was inspired by the idea that radiological images are a largely untapped resource for precision medicine. Hundreds of thousands of images are taken each year at every large hospital for routine clinical purposes. Images are digital, standardized, and of very high quality—in fact, most images contain information that that the human eye literally can’t see. There is an enormous amount of potentially valuable information being collected that is not being fully utilized for diagnostic or prognostic purposes.
“Early diagnosis of chronic disease is one of the holy grails of modern biomedicine.”
RG: What were you able to do with this study?
Palmer: In our recently published work, deep learning allowed us to explore the “hidden” features and patterns in CT images of the thorax that even expert humans are less able to decipher. We want to one day use this technology to predict the onset of chronic diseases such as diabetes, heart disease, and cancer before any symptoms are evident. As a proof-of-principle for our idea, we started off by looking at the much simpler outcome: 5-year mortality.
Ultimately, we hope to translate this new knowledge to improve patient outcomes. Early diagnosis of chronic disease is one of the holy grails of modern biomedicine.
RG: How did you test it?
Palmer: We compared the performance of deep learning against more traditional image analysis techniques currently used in radiology research, and found no significant difference between the approaches. This is despite the fact that the traditional techniques directly incorporate expert knowledge by measuring biomarkers, such as the extent of coronary artery calcification. By comparison, deep learning automatically learned to recognise markers of mortality without human guidance.
RG: How does the AI’s ability to predict patient lifespans compare to that of a human doctor?
Palmer: This is a hard question to answer. Outside of oncology, doctors do not routinely predict future lifespan in clinical settings. However, previous research has used clinical data such as demographic information and survey responses to predict mortality. It is difficult to compare our results with this work, as we controlled our analysis for factors other studies have found to be highly predictive, like age and sex. Allowing for these and other differences in study design, we see that our system has a similar accuracy as many previous attempts, which are typically between 65 and 75 percent accurate.
We fully expect that incorporating the strongly predictive and readily available clinical information such as age and sex will improve our results, and we intend to explore this in further research.
“This is far better than a human doctor could do, and we are now thinking about how best to use this information in a clinical setting.”
RG: At this point, can doctors do anything with the information AI produces to help patients?
Palmer: In short, not yet. Our current models, which have built upon the foundational work described in our recent Scientific Reports paper, are capable of predicting 5-year mortality in a subset of patients being imaged with a thoracic CT scan in a hospital setting with around 80 percent accuracy. This is far better than a human doctor could do, and we are now thinking about how best to use this information in a clinical setting.
RG: What’s next for this technology?
Palmer: There is certainly a lot of interest in predicting mortality in many fields besides oncology, such as cardiology, nephrology, and pulmonology. One future possibility could be the use of mortality predictions to assist with clinical decision making around high-cost or dangerous treatments such as transplants. We will soon be turning out attention to the use of radiologic images such as CT scans to predict the onset of chronic diseases.
We are continuing our research, using a dataset of thousands of cases. We intend to incorporate the clinical data routinely acquired with each scan and begin to predict other medical outcomes, like the development of strokes and heart attacks.
This sort of technology does not replicate what radiologists currently do, rather it provides an additional value to medical imaging studies. We envision that this approach could guide treatment decisions, potentiate preventative care, inform cohort selection for clinical research, and act as a more responsive biomarker for chronic diseases and aging.
Lyle J. Palmer is a genetic epidemiology specialist at the University of Adelaide's School of Public Health.