Dr House Goes Digital as IBM’s Watson Diagnoses Rare Diseases
October 24, 2016
Doctor House is going electronic. Medics charged with diagnosing the kind of extremely rare diseases that Hugh Laurie’s consultant routinely spots in TV drama House have found that artificial intelligence can do a similar job – but in seconds rather than days or weeks.
From December, doctors at the University Hospital of Marburg’s Centre for Undiagnosed and Rare Diseases (known as ZusE in German) will start using IBM’s Watson to speed up their diagnoses. In 2011, Watson famously won the gameshow Jeopardy! after being trained on 200 million pages of general knowledge data. Doctors are now training it on peer-reviewed rare disease literature to help them spot unusual illnesses.
In House, life-saving diagnostic decisions are often based on a deep probing of patients’ backgrounds. The show’s scripts are informed by authentic medical literature, and endocrinologist Jürgen Schäfer at ZusE says he once diagnosed one of his own patient’s debilitating fever – related to cobalt contamination from hip implants – after watching an episode of the show.
This got the ZusE team wondering if automating the analysis of patient information could help reduce their 6000-person waiting list. “Some of our very desperate patients have a very long medical history – some bring us five kilograms of paper records,” Schäfer says. But working with that information to reach a diagnosis can take days or weeks as experts pool their knowledge and deliberate in conference sessions.
To see if a computer could improve this process, Schäfer and his colleague Tobias Müller took 500 past patient records from their files and worked with IBM engineers to feed them into Watson in a format the system could understand. They then took a couple dozen sample cases at random and checked what Watson had made of them.
In all cases, Watson gave a series of potential symptom causes ranked by disease likelihood – and the top ones all coincided with what the doctors actually found at the time. But unlike those doctors Watson did the job in just seconds, says Müller.
He and his colleagues will now start using Watson as a diagnostic tool in new cases. To make it work, patients fill in exhaustive questionnaires to give Watson a view of their lives that might reveal potential disease exposures.
“It even asks them about their childhood and what pets they have,” says Müller. “We had one patient with inexplicable gut symptoms who, it turned out, kept an aquarium. He had caught the tropical disease bilharzia from his water snails.”
As their work with Watson progresses, the team hopes it may also help them spot diseases that are entirely new to medicine. If it has not been described before, says Müller, the machine will be stumped.
That is a potentially powerful capability, says Peter Bentley, a computer scientist at University College London who has applied machine learning techniques to medical diagnoses.
“If they can now trawl the literature properly with Watson, and still nothing matches, it might mean you have a new area for disease research – and one you need to look at fast,” he says. “It could help researchers respond faster to unknown diseases and make the most of research funding.”
Sign up for updates straight to your inbox.