Categories
AI research

#APaperADay – The Last Mile: Where Artificial Intelligence Meets Reality

“…implementation should be seen as an agile, iterative, and lightweight process of obtaining training data, developing
algorithms, and crafting these into tools and workflows.”

Coiera, E. (2019). The Last Mile: Where Artificial Intelligence Meets Reality. Journal of Medical Internet Research, 21(11), e16323. https://doi.org/10.2196/16323

A short article (2 pages of text) describing the challenges of building AI systems without understanding that technological solutions are only relevant when they solve real world problems that we care about, and when they are built within the systems that they will ultimately be used in.

Note: I found it hard not to just rewrite the whole paper because I really like the way Coiera writes and find that his economy with words makes it hard to cut things out i.e. I think that it’s all important text. I tried to address this by making my notes without looking at the original article, and then going back over the notes and rewriting them.


Technology shapes us as we shape it. Humans and machines form a sociotechnical system.

The application of technology should be shaped by the problem at hand and not the technology itself. But we see the opposite of this today, with companies building technologies that are then used to solve “problems” that no-one thought were problems. Most social media fits this description.

Technological innovations may create new classes of solution but it’s only in the real world that we see what problems are worth addressing and what solutions are most appropriate. Just because a technology is presented as a solution it’s up to us to make choices about whether the solution is the best solution, or whether the problem is important.

There are two broad research agendas for AI:

  1. The technical aspects of building machine intelligence.
  2. The application of machine intelligence to real world problems that we care about.

In our drive to accelerate progress in the first area, we may lose sight of the second. For example, even though image recognition is developing very quickly the use of image recognition systems has had little clinical impact to date. In some cases, it may even make clinical outcomes worse. For example when the overdiagnosis of a condition causes an increase in management (and associated costs and exposure to harm), even though treatment options remain unchanged.

There are three stages of development with data-driven technologies like AI-based systems:

  1. Data are acquired, labelled and cleaned.
  2. Building and testing technical performance in controlled environments.
  3. Algorithms are applied in real world contexts.

It’s only really in the last stage where it’s clear that “AI does nothing on its own” i.e. all technology is embedded in the sociotechnical systems mentioned earlier and are intricately connected to people and the choices that people make. This makes sociotechnical systems messy and complex, and therefore immune to the “solutions” touted by tecnology companies.

Some of the “last mile” challenges of AI implementation include:

  1. Measurement: We use standard metrics of AI performance to show improvement. But these metrics are often only useful in controlled experiments and are divorced from the practical realities of implementation in the clinical context.
  2. Generalisation and calibration: AI systems are trained on historical data and so future performance of the algorithm is dependent on how well the historical data matches the new context.
  3. Local context: The complexity of interacting variables within local contexts mean that any system will have to be fine-tuned to the organisation in which it is embedded. Organisations also change over time, meaning that the AI will need to be adjusted as well.

The author also provides possible solutions to these challenges.

Software development has moved from a linear process to an iterative model where systems are developed in situ through interaction with users in the real world. Google, Facebook, Amazon, etc. do this all the time by exposing small subsets of users to changes in the platform, and then measuring differences in engagement using metrics that the platforms care about (time spent on Facebook, or number of clicks on ads).

In healthcare we’ll need to build systems in which AI-based technologies are implemented, not as completed solutions, but with the understanding that they will need refinement and adaptation through iterative use in complex, local contexts. Ideally, they will be built within the systems they are going to be used in.


Note: I’m the Editor at OpenPhysio, an open-access, peer-reviewed online journal with a focus on physiotherapy education. If you’re doing interesting work in the classroom, even if you have no experience in publishing educational research, we’d like to help you share your stories.

By Michael Rowe

I'm a lecturer in the Department of Physiotherapy at the University of the Western Cape in Cape Town, South Africa. I'm interested in technology, education and healthcare and look for places where these things meet.