Deepmind’s AI works, but making money remains elusive

Deepmind plans more rigorous trials of its ophthalmology AI after early successes.

An artificial intelligence system from Google’s UK-based sister company Deepmind has beaten doctors in making referral decisions for patients with a range of retinal diseases, prompting talk of faster and cheaper diagnosis. The software will need rigorous controlled trials before it is market-ready, but in the meantime is to be provided to the UK’s NHS for free.

But Deepmind has been loss-making – increasingly so – since Google bought it in 2015, and speculation has been growing that the parent group is uneasy about how long Deepmind’s programmes might take to become profitable. 

The terms of Google’s acquisition of Deepmind were typically opaque, though the figure of £400m ($510m) has been reported. The group subsequently reorganised under a holding company, Alphabet, meaning Google and Deepmind are now on equal terms as Alphabet subsidiaries. 

Deepmind made a loss of £94m in 2016, the latest year for which data are available, compared with a £54m loss the prior year. Certainly Alphabet has money to burn, having ended last year with cash and assets of $124bn, so perhaps this is not a pressing problem, but eventually the parent is going to want Deepmind to demonstrate its value.


The recent pilot trial is a start. In partnership with Moorfields Eye Hospital in London, Deepmind’s AI was used to analyse optical coherence tomography (OCT) scans produced by two different imaging systems and make referral decisions for the patients. Patients had a range of disorders including central serous retinopathy, diabetic macular oedema and age-related macular degeneration.

After a training period using nearly 15,000 scans, Deepmind’s software made 55 incorrect referral decisions out of 997 using images from the first system, giving an error rate of 5.5%. It made four incorrect referral calls out of 116 using images from the second system, for an error rate of 3.4%.

Four senior ophthalmologists and four optometrists trained to interpret OCT images also looked at the images, and were matched or beaten by the system. The top two retinal specialists got 6.7% and 6.8% of their decisions wrong when they made them based solely on the scan results. When allowed access to patient notes and fundus images to inform their referral decisions, they could still only match, not beat, Deepmind’s system.

Deepmind is allowing Moorfields free access to the algorithm for non-commercial research for at least five years. Moorfields will retain control over the database, suggesting that Deepmind will have few or no rights to patients’ data. Clinical trials are to start in 2019.


Questions over Deepmind’s business model remain. According to an independent report commissioned by Deepmind and published in June, the company has no plans to sell patient data, even in a depersonalised form, and clear restrictions on how it can use data are stated in the contracts it signs with its partners.

The report’s authors go on to ask a very pertinent question: “If it is intended that Deepmind Health should make a profit, where will it come from?”

Dismissing the notion that Deepmind will monetise patients’ data somehow, this leaves two possibilities. The company could come to be a sort of loss-leader for Alphabet, never cash-generative but posting dazzling clinical results that lure clients to Google’s other AI businesses – whilst, almost incidentally, improving the speed and accuracy of diagnosis and cutting healthcare system costs.

Alternatively, the company really is ploughing the traditional furrow and developing a commercial product that could, if clinical trials succeed, be sold to hospitals. That will take time, and a lot of Alphabet’s money.

Share This Article