Deep-learning computers analyze patients’ risk of death with common body scans that could help doctors catch signs of diseases earlier.
Every year there are 85 million CT scans taken in the United States. Now researchers in Australia say that data from these body scans could be used to predict a person’s risk of death in the next five years — and perhaps prompt them to make changes that could extend their life.
Researchers have previously used genetic and environmental data — from diet to exercise habits — to estimate the life span of individuals, but a study from the University of Adelaide was the first to use artificial intelligence (AI) to analyze patient CT scans to predict mortality.
“Such knowledge is critical for improved early intervention, for better treatment decisions, and for ameliorating the steadily worsening epidemic of chronic disease,” according to the study published in the journal Scientific Reports.
The researchers used a machine learning program to analyze 48 chest scans from adults over 60 years of age. And a computer accurately predicted the chance of death within five years 69 percent of the time. The team used old CT scans, along with data on whether the patient died within five years, or lived longer, which they used to verify the computer’s prediction.
The 69 percent accuracy rate is comparable to genetic and environmental-based forecasting of future life span, according to researcher Dr. Luke Oakden-Rayner, a radiologist at Royal Adelaide Hospital and a PhD student.
The results show that AI analysis of CT scans could be a powerful and efficient new tool for doctors when analyzing patients’ health and risk of death — or catching harmful conditions in an earlier stage.
This information could be used to guide healthcare choices, including preventative measures, such as lifestyle changes. Similar information is already provided to patients with chronic diseases such as cancer, and many people do find it motivating.
“We hope that systems like ours will be able to provide this information earlier, when there is more opportunity to prevent serious complications,” said study co-author, Lyle J. Palmer, PhD, professor of epidemiology at the University of Adelaide.
The small size of the study left researchers unable to determine the basis of the computer’s predictions, but a follow-up will apply the same model to tens of thousands of CT images.
“Extensive data is crucial for training deep-learning systems to make accurate predictions,” said Robert Hudyma, MSc, associate professor at Ryerson University’s School of Information Technology Management.
Hudyma, who isn’t associated with the study, noted that a company called Enlitic trained its AI system on more than 17,000 chest X-rays to “diagnose cancerous tumors with a greater accuracy than a radiologist or oncologist can.”
“What Enlitic is doing is improving tumor diagnosis, which has the potential to save lives,” he said. “Similar work is underway to take a picture of a skin lesion with a smartphone and essentially diagnose various skin conditions using a neural network instead of a dermatologist.”
As with genetic information, the use of AI to predict future health problems and life span raises ethical concerns, including how such data could be misused by employers and insurers.
However, there are major potential benefits of using AI to catch diseases early, Palmer said. “Prevention is almost always cheaper than a cure, so insurers and employers will have a financial interest in funding measures that keep individuals healthy.
“While there are legitimate concerns about misuse of private medical data, there are no ethical issues specific to the data we are generating,” said Palmer. “And we think that the sort of approach we motivate will have benefits that far outweigh any privacy-related risk.”
But don’t expect routine annual CT scans anytime soon. A recent analysis found that exposure to radiation from multiple unnecessary body scans significantly increases risk of cancer. However, a single scan is generally considered safe.