Bing web searches may reveal you have cancer (so, er, don't use Bing?)

Microsoft eggheads crunch queries for early tell-tale signs of disease

24 Reg comments Got Tips?

Search engine results can be a useful predictor for cancer and can even beat doctors to the mark, according to new research from Microsoft.

In a paper published in the Journal of Oncology Practice, a Redmond team detailed how, by analyzing data from Bing users who looked up symptoms of pancreatic cancer, they could get an accurate diagnosis between 6 and 32 per cent of the time, with an error rate of 0.00001 per cent.

"We have shown that we can predict evidence of the future issuance of experiential queries about pancreatic cancer well in advance of their appearance in individuals' query streams at low error rate," the paper states.

"The success of these methods has implications for online methods that would provide passive screening of searchers with a view to providing early warning about potential signs of pancreatic adenocarcinoma and other devastating diseases."

The team built its data set using 9.2 million queries from US users searching for illnesses that match the early symptoms of pancreatic cancer.

The Microsofties looked for users who had been diagnosed with the cancer, and then worked backwards through their search history to see what had been requested before they were diagnosed. In other words, the team looked for clues in queries that suggest someone has cancer before they know they have the disease.

Therefore, if someone else has a similar pattern of searches, they have similar symptoms and may have pancreatic cancer.

All the search data was, of course, anonymized, so the team surmised a diagnosis from user searches for specific treatments, such as pancreaticoduodenectomy, or for cancer medications such as gemcitabine and 5-fluorouracil. Also included were obvious indicators, like searches for "i was told i have pancreatic cancer what to expect" and "just diagnosed with pancreatic cancer."

They then searched through the past history of searches looking for times when people binged common symptoms of pancreatic cancer, such as sudden weight loss, dark urine, or blood clots.

By refining down the algorithm, researchers found they could spot a buildup of threatening symptoms coming in and thus get an early prediction. Such an early warning could be crucial in the case of pancreatic cancer, which is hard to spot and kills 75 per cent of sufferers in the first year.

However, there's a lot more research to be done to improve the rate of detection before this can be rolled out. Mind you, given the number of hypochondriacs online, they probably wouldn't have a shortage of people signing up for trials. ®


Biting the hand that feeds IT © 1998–2020