Want to know if you’ll be dead in five years? Just let a computer look at your organs.
New research has indicated that “future” predicting computers could be coming to hospitals in the near future. Researchers are hoping that the technology could be used to predict serious illnesses and medical conditions such as heart attacks.
For the study, five year–old medical images of 48 patient’s chests were analyzed by artificial intelligence. From these images alone, the system was able to predict (with 69 percent accuracy) whether or not a patient would die within five years. It was also able to predict medical outcomes by analyzing large volumes of data and discovering subtle patterns. This new exam–by–AI has proven to be more effective than a physical exam from a doctor, though apparently it’s really not a fair fight.
“Human doctors are not trained to predict mortality, so the comparison is a bit unfair,” study leader Dr. Luke Oakden–Rayner of the University of Adelaide told Fox News. Oakden–Rayner added that previous research using clincial data such as age, sex or physical fitness had between 65 percent and 75 percent accuracy, so the new study "compare[s] favorably, especially considering we excluded factors like age and sex from our analysis."
Currently the system is only trained to predict death within five years, although, according to Oakden–Rayner, with the right dataset it should be trivial to extend the technique to other time scales.
To predict mortality at ten years, for example, the system would need to analyze CT scans performed over ten years ago so that Oakden–Rayner’s team could have the follow up results. The researchers also couldn’t tell what exactly it was the computers were seeing in the images to make their assessment, though they did find a strong relationship between the prediction of mortality and the presence of visible illnesses such as emphysema and congestive heart failure.
Oakden–Rayner explained that certain techniques could be applied to visualize how the computers were seeing the scans, but the study was too small to use them this go–round.
“You can identify the regions of the images that contributed to the prediction, and you can ‘hallucinate’ images that exaggerate the features that the system uses (generating exemplars of "survival" and "mortality" scans),” he said. “We couldn't apply these techniques effectively due to our small dataset, but are currently applying them to a much larger group consisting of tens of thousands of patients.”
For this next round of testing, Oakden–Rayner and his team will be incorporating highly predictive clinical information like age and sex into their models, which they expect will improve prediction accuracy.
Similar medical AI news has been cropping up lately: a startup in China revealed an AI system that can help doctors identify lung cancer by examining CT scans, and IBM now has AI in hospitals (called Watson) that can answers patient questions. However, if perfected, this new medical AI could be the most exciting development yet.
“We will start looking at predicting other medical events before they happen, like strokes, cancer and heart attacks,” Oakden–Rayner said.
The study can be found in "Scientific Reports."