'Virtually no difference' between AI and humans in diagnosing prediabetes
Is that... a good or bad thing?
Deep-learning algorithms have shown themselves equal to humans in detecting patients at high-risk of developing Type-2 diabetes by analyzing CT scans of their pancreases, according to a research paper published on Tuesday.
Type-2 diabetes is estimated to affect 11.3 percent of the US population, or at least 37 million people. Type-2 diabetes can lead to issues with circulatory, nervous, and immune systems, increasing the risk of heart disease and strokes.
Those with the initial form, prediabetes, can repair their body's insulin resistance, so they don't develop the full-blown condition, if they change their diets and exercise habits. American health officials reckon 38 percent of the US adult population, some 96 million people, have prediabetes.
Now, a team of researchers has developed a new method using an AI model to automatically detect prediabetic patients, and the results show "virtually no difference" between the accuracy of the AI's forecast and human work.
"The analysis of both pancreatic and extra-pancreatic features is a novel approach and has not been shown in previous work to our knowledge," said Hima Tallam, first author of the paper and a PhD student at the US government-funded National Institutes of Health (NIH).
The model based on a convolutional neural network looks at the density and fat content in the pancreas to determine whether a patient has early onset diabetes or not.
"We found that diabetes was associated with the amount of fat within the pancreas and inside the patients' abdomens," Ronald Summers, co-author of the study and a staff radiologist at the NIH said. "The more fat in those two locations, the more likely the patients were to have diabetes for a longer period of time."
- DARPA to build life-saving AI models that think like medics
- Google helps develop AI-driven lab machine to diagnose Parkinson's
- AI-designed drug to treat deadly disease now tested on humans
- Machine learning the hard way: IBM Watson's fatal misdiagnosis
The team trained the proof of concept model on a small experiment with 471 images from three different datasets, eight images were used for validation and 39 for testing. The system was tested further on 25 patients randomly selected from a group of 8,992 people, where 572 of them had been diagnosed with Type 2 diabetes, and 1,880 had dysglycemia, a medical condition that makes blood sugar levels too high or low associated with prediabetes.
A radiologist was given the same images from the randomly-selected patients and the results were compared against the neural network model. The automated methods performed just as well as the human expert, the researchers claimed. They improved the software further by adding more data such as a patient's BMI.
The team believes AI can diagnose prediabetic patients faster than health workers can and prevent more people from developing Type 2 diabetes. "This study is a step towards the wider use of automated methods to address clinical challenges," the authors concluded. "It may also inform future work investigating the reason for pancreatic changes that occur in patients with diabetes." ®