The algorithms that are already changing your life

At Moorfields Eye Hospital in London, consultants are facing a familiar problem.

Patient numbers are surging. Age-related eye diseases are becoming more and more common, and as the British demographic gets ever older, numbers are predicted to increase by between a third and one half.

“We have enormous numbers of patients, we can barely cope,” says Professor Peng Tee Khaw, a consultant ophthalmic surgeon. “We need to look at new ways to deal with the issue.”

When a patient arrives at Moorfields, doctors will likely perform an eye scan that captures a 3D cross-section of the person’s retina. The images are complex and beautiful, but often hide subtle signs of eye disease. It takes an experienced pathologist to spot abnormalities and decide what treatment is needed and how urgently.

The process is not a fast one, but artificial intelligence is about to change that. Working with Google’s artificial intelligence group, DeepMind, doctors at Moorfields have trained an AI on a million anonymised eye scans from patients at various stages of age-related macular degeneration. The hope is that the AI will learn to spot the earliest signs of disease and ultimately deliver a diagnosis.

“If this is as accurate as a human being, the whole process of diagnosing disease and understanding what needs to be done can be done pretty well instantly,” says Khaw. It could make an enormous difference to Moorfields patients: for some conditions, early treatment can be sight-saving. The results so far are promising and a formal clinical trial could start as early as next year, Khaw says.

The Moorfields project is just one of a slew of instances where AI is making an impact. The technology is being woven into the fabric of life, to help people communicate, travel, meet partners and get loans. It targets customers to drive sales and monitors employees for suspicious behaviour. At the same time, it helps the emergency services, social workers and urban planners. For all its potential benefits though, critics warn that the rapid proliferation of such a powerful technology poses fresh threats to basic human rights, privacy and society in general. “There are certain standards that need to be in place for this to work well,” says Craig Fagan, policy director at Tim Berners-Lee’s Web Foundation. “Companies have to make sure that what they’re putting out is not creating social harm.”

Medicine is primed to be a chief beneficiary of artificial intelligence. AI can diagnose diseases from telltale groups of symptoms, strange patterns in blood tests, and the subtle abnormalities that cells display as a disease begins takes hold. Time and again, AI systems are found to pick up signs of illness that are unknown to doctors, making the AIs more accurate as a result. Earlier this year, researchers at Nottingham University trained several AIs to spot people at risk of heart attack and found that all of them performed better than doctors.

Another AI built at Stanford University in California has learned to spot breast cancer in biopsy tissues. Pathologists typically make the diagnosis after checking a handful of tissue features, but the AI outperformed the cancer specialists by considering more than 6,000 factors.

Researchers have begun to use AI in mental health too. A Boston-based company, Cogito, is trialling a mobile phone app that monitors the tone of a person’s voice to detect mood changes that could flag a bout of depression. In China, researchers want to spot those at risk of suicide from their posts on Weibo, a Twitter-style microblogging platform.

Treatment is also ripe for an AI-fuelled revolution. Algorithms trained on piles of medical records can advise doctors on the most effective drugs for the patient before them, taking into account their genetic makeup and other conditions they have. Its success now relies as much on finding effective ways to share patients’ medical data without putting privacy at risk.

A UK government review of AI in October proposed “data trusts” that would allow the NHS, for example, to share sensitive information securely. Done well, the trusts could potentially prevent more unlawful uses of data, as happened when the Royal Free Hospital in London shared the health records of 1.6 million identifiable patients with DeepMind for its own artificial intelligence project.

Toby Walsh, professor of artificial intelligence at the University of New South Wales, and author of a recent book on AI called Android Dreams, fears that a small number of tech giants could come to own our health and other data, giving them enormous power over our lives. “It will look like 1984, but it won’t be a government that’s in charge, it’ll be a corporation, and corporations are even less answerable than governments,” Walsh said. “In 10 or 20 years time, if Google is not broken up into separate parts, I will be severely worried for the future of democracy.”

The boom in AI applications will reach far beyond medicine. Online retailers have ramped up the use of AI to maximise sales; some dating sites use the technology to match potential partners; and cities such as Manchester are dabbling with AI-controlled traffic lights to ease congestion and reduce air pollution.

AI-powered cyber defences have also arrived. The UK-based company Darktrace uses AI to spot suspect activity on companies’ computer networks, a strategy that revealed the curious case of a North American casino that was hacked from Finland via its wifi-controlled fishtank. Darktrace recently detected a worrying new form attack: while monitoring activity for an Indian company, the tech firm spotted AI-enhanced malware that learned how to blend into its target network and lurk there without detection. Since India is one of the world’s testing grounds for new cyber attacks, more AI-powered malware could soon be targeting companies around the world.

AI is already helping the police to tackle crime. In 2014, a Kent police officer was on his way to interview the victim of a double motorbike theft when he heard the meeting had been delayed. With an hour to kill, the officer went to a nearby area that had been flagged that morning as ripe for crime by PredPol, the force’s AI tool. During the officer’s patrol, he spotted the missing motorbikes, made an arrest, and had the bikes returned to their owner.

Kent police has pioneered predictive policing in Britain. Having trialled and adopted PredPol, a US commercial product, in 2013, the force has gained more experience than most. Sceptical at first, officers introduced the tool after a trial revealed PredPol was 60% better at spotting where where crimes would take place than the force’s analysts. “There was nothing we could do that was more accurate,” said Jon Sutton, head of transformation, performance and analysis at Kent.

View image in fullscreen

PredPol uses artificial intelligence to learn crime patterns from historical records. The Kent system was trained on five years of crime data, and the algorithm is now updated daily with the force’s most recent three years of records. After crunching the data, PredPol returns a daily list of 180 hotspots, each 500 foot by 500 foot, where it predicts the crime risk is high. About 80% of the boxes never change: some areas always attract more crime than others. But the rest move around in line with patterns PredPol has learned from years of criminal activity. Some patterns are obvious and follow the clock or the seasons. There are more brawls near pubs and clubs at night time, and more incidents around Kent’s beaches in the height of summer. Others are more subtle and reflect trends in crimes, the movements of gangs, or new vulnerabilities in particular neighbourhoods.

The 180 PredPol hotspots cover about 0.1% of Kent, but within them, about 17% of crime and 21% of antisocial behaviour takes place. Officers are not sent to cover all the hotspots. Instead, the police, along with community support officers and community wardens, are briefed on the locations and conduct visible patrols in the area when they can. “It’s one of a number of crime prevention tools, but our officers have made arrests in areas where they say they wouldn’t have been were it not for PredPol,” said Sutton. “We don’t see it as a panacea. It doesn’t replace skills, knowledge and experience.”

Predictive policing has its critics though. A recent study by the University of Utah found that the software could trigger “runaway feedback loops” where officers are sent back to the same, often poor, neighbourhoods time and again. The problem arises when police in a hotspot make an arrest, leading the software to rank the area as an even higher crime risk area in future, and so send more police back the next day, regardless of the true crime rate.

In Kent, the police, community support officers and wardens only patrol PredPol boxes between scheduled duties. With 70% of their patrols having no power of arrest, the tool is primarily used to prevent crime rather than catch criminals. One patrol, for example, noticed a line of industrial bins lined up beneath open windows on a housing estate. “It was just a case of putting the bins on the other side of the car park,” said Nicola Endacott, Kent’s deputy head of analysis.

According to the government’s October AI review, the rise of AI has brought us to the threshold of a new era, with profound implications for society and the economy.

“Quality of life might very well be improved. In terms of solving the big problems from climate change to the supply of energy, AI should be able to help,” said Dame Wendy Hall, a co-author on the report.

“It’s going to be big.”