To apply this technology to an assisted-living facility and accomplish the same things that the single-patient tool is meant to do, you would need to build completely new A.I. models from scratch, starting with acquiring and curating a new set of training data.
Why? Because A.I. models are often incredibly sensitive to the training data they use. If you only use training data that consist of daytime images, the model will produce errant results on images taken at night. If you use data from one hospital, it could produce errant results for a different hospital. And if you use images from one camera, it could produce errant results with images taken using a different camera.
This may seem strange—images from one camera are fairly similar to images from another camera. But since the training data provide all the information the model has to produce results, it uses signal from anywhere it can, and that could be something seemingly minute, such as the field of view or the color density of an image. This sensitivity of models complicates every discussion about adapting a product for an adjacent opportunity.
To the business strategist, expanding into assisted living seemed obvious; from a tech standpoint, it would have required a full rebuild. Leaders of A.I. companies need to think about both the market opportunity and the technical implications of a new idea before deciding which path forward makes sense.
Don’t forget the workflow
Traditional strategy emphasizes the importance of making sure the product fits within the customer workflow. There are lots of examples of this—gyms, for instance, tend to want locations near offices because most people work out before or after work or during their lunch break. Setting up a gym far from office buildings wouldn’t take into account the customer workflow.
Similarly, LookDeep has had to make sure its product integrates with the daily routines of doctors and other healthcare providers. But with A.I., considering the workflow means confronting another data-acquisition challenge. For example, if a hospital using LookDeep wants to alert a nurse every time a patient has been lying still in bed for more than two hours (and is therefore at risk of a pressure injury), it needs to consider the nurses’ workflow to make sure the product isn’t disruptive to their day-to-day operations. A nurse may be on break, not on shift, in a different part of the building, or with another patient, or there could be an altogether better person to alert based on the nurses’ workflow. Simply detecting when a safety check is warranted is just the first hurdle to clear; considering all parts of the operational context so that the right person is alerted at the right time and in the right manner may be as challenging or even more so.
As a result, for LookDeep to provide this functionality without impacting workflow, it would need identifying data not only about when patients are static and in bed, but also where their rooms are located and who the nurses are they’re matched with, as well as all nurse staffing and shift records, just for starters.
The nuance with A.I. is that even though it is still important to think about how a solution works for the customer, your data-acquisition goals change based on the customer workflow. This means you have to identify how the solution aligns with the workflow up front when deciding what functionality to offer and how to build the training data for your algorithm.
The code isn’t the key
We started out by saying that strategy and technology are inextricable for A.I. companies. That may sound like a problem, as researchers continue to move the science of A.I. forward at a rapid pace. But aligning strategy with technology doesn’t mean shaking things up with every new innovation in computer science and machine-learning methodology. The aspect of A.I. that corporate leaders need to keep in focus is not so much the code as the data.
When LookDeep decided to build a posture detector (to determine whether patients are sitting upright), for example, the company’s team spent most of its time on the project defining the data and the problem. They then selected from several options for classification models—they did not create a model from scratch.
What differentiates LookDeep’s product is the hours spent building a data set and ensuring its robustness (e.g., testing it with multiple people in a room, at different times of day). This arduous approach to data generation not only increases the accuracy of the algorithm dramatically but also serves as a strategic and competitive advantage. Anyone who wants to compete with LookDeep’s product will have to make the same investment in data.
Singh’s decision to invest more in data than in code underscores the notion that successful executives at A.I.-oriented companies don’t need an advanced understanding of the technical minutiae but rather a working knowledge of the principles of A.I. and the way they affect strategic decision-making. As industries continue to find new means by which A.I. can improve operations and lower costs, there is a need for businesspeople who understand the process of building an algorithm and know what questions to ask. Without those people to connect technology to strategy, companies can end up making expensive mistakes as they try to integrate A.I. into their business models.
Naila Dharani is principal consultant for Chicago Booth’s Center for Applied Artificial Intelligence.
Jens Ludwig is the Edwin A. and Betty L. Bergman Distinguished Service Professor at the University of Chicago Harris School of Public Policy.
Sendhil Mullainathan is the Roman Family University Professor of Computation and Behavioral Science at Chicago Booth.