Summary: Researchers report that artificial intelligence systems can predict a person’s lifespan by analysing CT images of their organs.
Source: University of Adelaide.
Artificial intelligence that estimates a patient’s lifespan by examining standard medical scans has moved closer to clinical reality, according to a new study led by the University of Adelaide.
The study, published in the journal Scientific Reports, demonstrates that computer-based analysis of chest CT images can predict five-year mortality with accuracy comparable to clinicians’ manual assessments. These findings have potential implications for early diagnosis, risk stratification and targeted medical intervention.
Researchers from the University of Adelaide’s School of Public Health and School of Computer Science, together with Australian and international collaborators, applied deep learning and radiomics techniques to routine CT imaging from 48 patients. The algorithm correctly predicted which patients would die within five years with 69% accuracy, a level similar to that achieved by human experts using clinical judgment.
This work represents one of the first demonstrations of machine learning applied to medical imaging specifically for longevity prediction rather than solely for disease diagnosis.
“Being able to predict a patient’s future health trajectory allows clinicians to personalise care and focus interventions where they are most needed,” says lead author Dr Luke Oakden-Rayner, a radiologist and PhD candidate at the University of Adelaide’s School of Public Health. “Current assessments of biological age and overall organ health are limited because clinicians cannot easily quantify the combined burden of disease inside the body from routine imaging alone.”
The team used deep learning, a branch of artificial intelligence in which neural networks learn image features automatically, and radiomics, which extracts quantitative features from medical images. Although the study used a modest sample size, the results suggest the computer system learned to recognise complex imaging patterns associated with poor outcomes—patterns that can be challenging and time-consuming for humans to detect consistently.
In this pilot study the computer model could not be reduced to a simple list of visible signs that explain every prediction. However, the most confident predictions correlated with patients who had severe chronic conditions, including emphysema and congestive heart failure, indicating the algorithm captured meaningful markers of advanced disease.
“Rather than simply diagnosing individual diseases, these automated systems can synthesise large volumes of image-derived information and detect subtle, distributed patterns that contribute to overall prognosis,” says Dr Oakden-Rayner. “That capability complements clinical judgment and may open new pathways for precision radiology and preventive medicine.”
The authors emphasise that this research is a proof of concept rather than a ready-to-deploy clinical tool. Future work will need to validate these methods on much larger and more diverse patient cohorts, integrate clinical and demographic data, and improve interpretability so clinicians can understand which image features drive risk predictions.
Next steps announced by the research team include scaling the approach to tens of thousands of patient images to improve model robustness and generalisability. They also plan to investigate whether similar techniques can predict specific outcomes such as the risk of heart attack or other major events, rather than overall mortality alone.
By applying convolutional neural networks and feature engineering within a radiomics framework, the study illustrates how routinely collected CT scans could yield prognostic biomarkers for precision medicine. Automated image analysis may therefore become a complementary tool to support earlier intervention and better-targeted therapies for patients at high risk of adverse outcomes.
Source: David Ellis – University of Adelaide
Original Research: Full open access research titled “Precision Radiology: Predicting longevity using feature engineering and deep learning methods in a radiomics framework” by Luke Oakden-Rayner, Gustavo Carneiro, Taryn Bessen, Jacinto C. Nascimento, Andrew P. Bradley & Lyle J. Palmer in Scientific Reports. Published online May 10, 2017. doi:10.1038/s41598-017-01931-w
University of Adelaide. “Artificial Intelligence Predicts Patient Lifespans.” NeuroscienceNews. June 2017.
Abstract
Precision Radiology: Predicting longevity using feature engineering and deep learning methods in a radiomics framework
Precision medicine depends on accurate, non-invasive assessment of an individual’s true health state, reflecting genetic risk and environmental exposure. Currently, there is a lack of efficient tests to capture the full range of phenotypic variation relevant to individual health. Such information is essential for earlier intervention, better treatment decisions, and to address the growing burden of chronic disease. This proof-of-concept study demonstrates that routinely acquired cross-sectional CT imaging can be analysed with computer image analysis methods to predict patient longevity as a proxy of overall health. Despite a modest dataset and the use of standard machine learning methods, results were comparable to previous manual clinical approaches for longevity prediction. The work shows that radiomics can extract prognostic biomarkers from medical images and that deep learning with convolutional neural networks can be applied productively in radiomics research. Applying computer image analysis to routinely collected medical scans offers significant potential to enhance precision medicine initiatives.
“Precision Radiology: Predicting longevity using feature engineering and deep learning methods in a radiomics framework” by Luke Oakden-Rayner et al., Scientific Reports. Published online May 10, 2017. doi:10.1038/s41598-017-01931-w