There’s never a good time to get bad news from your doctor. But for some diseases, early diagnosis is crucial. ERC-funded researchers are studying ways to apply computer power to identifying and treating disease
By Gary Finnegan
For centuries, we have relied on the expertise of doctors to examine patients, consider their medical history, and make a diagnosis. In recent decades, new blood tests and imaging technologies have added extra information into the mix. Next up: The digital doc.
Nikos Paragios, a computer science and applied mathematics professor at France’s École Centrale des Arts et Manufacture, leads a project funded by the European Research Council that is trying to tap the power of artificial intelligence (AI) to build models that predict how patients’ organs or tumours will develop in the years ahead.
“The best way to optimise treatment outcomes is by taking all the available information into account,” he says. By feeding large volumes of information into his computer model, Paragios says, an algorithm – like the mathematical formula that powers Google or allows Amazon to predict what books you might like to buy – can help doctors build a tailored treatment plan for their patient.
“We want to move from generic disease treatment to a more personalised digital medicine,” he says. “The ideal system would be able to predict outcomes earlier than is currently possible.”
Paragios is working in a growing field of research around the world. At Stanford University, a research group this year reported progress in using AI to diagnose skin cancer. In China, another group report using AI to manage congenital eye cataracts. Several companies have been piling in, as well – most notably, IBM with its much-publicised Watson supercomputer. And governments are responding: for instance, in the US, the Obama administration included AI diagnostics in two of its flagship health-research programmes, the ‘Cancer Moonshot’ and the Precision Medicine Initiative.
AI for Alzheimer’s
And in Europe, another ERC-funded researcher, Stanley Durrleman of the INRIA/ICM AramisLab at the Brain and Spine Institute (ICM) in Paris, is working to predict the development of Alzheimer’s Disease in individual patients. His aim is to compile brain scans from diagnosed patients, and then use that to better understand how the brain works and what it looks like when things go awry.
“Diagnosis of Alzheimer’s disease comes very late,” Durrleman says. “Changes to the brain probably begin much earlier that can currently be detected; but we know very little about this silent stage of the disease.”
Durrleman’s work combines mathematics with years’ worth of brain scans. “We want to use digital models to predict how many years it will take for symptoms to develop,” he says. “This will allow us to make a personalised prediction for individual patients.”
Cancer patients are another potential beneficiary of this technology. For them, Paragios says, the patient’s age and information on biomarkers and body scans can be checked against a database that shows how other patients fared when treated with various combinations of radiotherapy or mediation.
“In around 25 per cent of cases, treatment plans for cancer patients should be adjusted because of changing anatomy,” Paragios says. “That’s not happening at present because we are not making full use of the data. AI could change that.”
Humans need not apply?
For some tasks currently performed by humans, machines offer greater speed and accuracy. How close are we to handing over to digital doctors?
“Algorithms can be better than humans but they need a lot of data which has been well annotated,” Paragios says. “There are some computer-aided diagnosis tools on the market now, but in the next five to 10 years, we expect to see many more.”
Today, computers already perform a small number of complex tasks – such as calculating the dose of radiation for cancer patients. More diagnostic and treatment decisions will gradually shift away from established professions such as radiology.
“One of the challenges is to convince radiologists to buy software that will replace them in the long run,” says Paragios. “It’s not that radiology will disappear; but there will be a shift to interventionist radiology with greater automation of specific tasks which are currently time-consuming or tedious.”
Of course, the machines are only as good as the data they are fed and are best-suited to answering very specific questions. Paragios says the software currently being developed is not ‘reasoning’ in the way humans do, but is capable of scanning huge databanks, crunching the numbers and advising health professionals what to do next.
This raises a couple of big challenges. For a start, access to large volumes of high-quality, relevant data is essential. The legal and ethical issues surrounding ownership of this information is a major challenge.
A second problem arises in getting certification for technologies that are making life-changing ‘decisions’. Even if the software is more accurate than humans, who is responsible if the computer misses a bunch of malignant cancer cells?
“At present, the tools available just make a recommendation and it is up to the physician or radiologist to make the call,” Paragios says. “Looking ahead, the new class of digital tools will be making the call. That’s what we have in radiation therapy where the computer calculates the dose. In future, we’ll be working primarily with this class of system.”
Paragios and Durrleman were among ERC grantees speaking at a debate on longevity at BioVision in Lyon on 5 April.