AI is being pioneered in medical institutions across the world; however, two questions need to be answered before its usage can be justified further: is it effective, and what are the risks involved? This article will consider this in relation to a recent effort to use AI to help treat eye diseases.

Google’s DeepMind subsidiary, UCL, and Moorfields eye hospital have used deep learning to develop software that identifies dozens of common eye diseases from 3D scans and then recommends the patient for the appropriate treatment, the software is sensitive to more than 50 eye conditions.

The research has been described as ‘ground breaking’, with Mustafa Suleyman, head of DeepMind Health, saying this technology could be used ‘around the world’ to prevent life threatening illnesses.

The software itself is based on established principles of deep learning, which uses algorithms to identify common patterns in data. The relevant data in this study being the 3D scans of patients eyes made using a technique referred to as optical coherence tomography.

It takes ten minutes to create one of these scans- this is achieved by bouncing near-infrared light off the interior surfaces of the eye.

Learning took place by feeding the system a series of scans alongside diagnoses made by human doctors; from this it learnt to identify the different anatomical elements of the eye and recommend clinical action.

The AI was tested versus a panel of eight doctors and made the same recommendation 94% of the time.

That said, these results do engender some concern. Many, for example Luke Oakden-Rayner a radiologist who has written extensively on AI, feels a tipping point is approaching where software is no longer applied and interpreted by a doctor, rather it makes decisions on the behalf of doctors; the problem being that there is nothing to scrutinise these judgements, potentially endangering the patient.

However, it would appear these criticisms have been dealt with by the research team. For example, they have mitigated the risk of an improper diagnosis by basing the software on multiple algorithms instead of one; thus, any freak error made by a single algorithm will become overruled by the majority. Furthermore, it is intended to be used for triage- i.e. deciding which patients need attention first- thus while it guesses what a patient might have, human medics will be on hand to confirm this to further prevent a false diagnosis.

This is another example of the potential benefit algorithms and deep learning could bring to us, this research will hopefully have a marked impact on the lives of those it is used on.

Furthermore, this demonstrates how it is possible to prevent many of the criticisms often thrown at AI, and that- if proper research design is carried out- we can minimise risk whilst maximising our return- in this case, that return being one of our most treasured experiences: sight.