I was recently reading an article about Vertical Farming – crops are grown indoors in a ‘stack’ under controlled environmental conditions (i.e. lighting) and using no soil. (If you want to read the book that bought it to the mainstream, and I have just started to read it, check out The Vertical Farm: Feeding the World in the 21st Century).
Anyway, the article highlighted the work of a leading start-up in the space called Infarm. They ship, what was described as ‘vending machines’, to clients (including Aldi, Carrefour and Marks & Spencer), where the greens and herbs grow using LED lights and nutrient rich water. The founders had done countless experiments on lighting, nutrients, humidity, etc to get the right ‘mix’ for the greens and herbs to taste just right. You can just pick the plant you want straight from the machine it grows on. Cool, eh?
What was more interesting was what Infarm are planning next; larger scale units that are super-efficient promising much higher yields. This was driven by the fact that the amount of manual work required and the low price of the product didn’t make for a great business model (even start-ups have to make money at some point). They have combined automation (robots that adjust the position of the plants inside a farm including the light they get) and machine learning (to optimise the ‘variables’ of lighting, nutrients, humidity, etc to get the perfectly crispy lettuce). Their ultimate plan is to build a network of tens of thousands of these automated farms which will feed a central ‘brain’ (or master algorithm) that will in turn come up with more efficient ways for individual farms to grow their crops.
What does this have to do with health?
Well, the distributed / centralised machine learning setup made me think about federated learning – where an ‘end point’ would take a ‘copy’ of the global model, improve it using the local data it has available and only share back with the ‘centre’ incremental improvements to the global model.
Why is this important for health?
It has been shown that many of the algorithms that were developed to predict Covid do not work and that some were even harmful. Poor quality training data was a key issue. Other issues were AI technologists and clinicians working in siloes, and incorporation bias (i.e. incorrect labelling based individual clinician opinions). It was additionally highlighted that most researchers developed their own models rather than collaborate or look at how their work could improve existing models. Sharing data was suggested as a potential solution, but that always has a myriad of data privacy challenges to overcome.
So, imagine a scenario where public sector funding is provided to develop a centralised model (and data to do initial training of the model) and distributed models operating at individual hospital / health ecosystem level. Innovators will focus on the distributed ‘end points’, applying their specialist knowledge of the use case, designing systems in a human centred manner that engages and implementing scalable systems into health (with the inherent change management challenges). The ‘centre’ is managed by a public / public-private (independent to ‘end point’ providers) / public – academic / public – academic / private partnership. Overall, each ‘end point’ feedbacks into the central model (i.e. we learn from the individual implementations) and in turn, each ‘end point’ benefits from the collective intelligence of the central model.
[PS: The other interesting element of the Infarm setup is the ‘onsite’ part, clients tell Infarm the products they want and they are physically on the client site. There are some parallels to the data privacy element that makes federated learning attractive in areas such as health. The benefits of data privacy would be preserved in the scenario described above].
Very happy to hear your comments below or feel free to email me to share ideas – firstname.lastname@example.org