
I noticed that a paper from the Moorfields/Google project was published over last weekend.
My first response from the general media reports was one of ‘here we go again.’ Statements such as ‘doctors harnessed machine learning – a branch of artificial intelligence (AI) dedicated to training computers to become smart’ are always a little disappointing. We have been using machine learning for years. Any instrument that includes a database, such as your fields screener or OCT, includes data previously input from a range of patient results and allows the machine to flag up anything suspicious (such as the p values, total deviation or GHT on a field plot, or the colour coding on most OCT outputs) is using machine learning. There is no ‘intelligence’ here beyond data storage and identification of outliers within a set range.
It is a shame when such flowery language is used to report such papers, as one possible result is that people assume the intelligence ascribed to machines will, in some way, replace that of a human operative. This is clearly not the case, and machine learning should be embraced as a way of assisting the clinician when interpreting any particular set of results.
Further reading of the new paper, however, shows that it actually concludes something new. The ease with which practitioners may enter their own data into the ‘machine’, in this case the automated deep learning platform Google Cloud AutoML, shows that clinicians and not IT specialists might soon be in a position to tailor algorithms to meet their own clinical needs, whether screening for a specific disease or profile pattern. Now this is interesting and may one day offer useful flexibility to eye care professionals working in specific fields, allowing them to customise any automated screening to best meet their own requirements.
Worth downloading and having a read.