Lirio reposted this
It was available online a few weeks ago, but I was waiting until the beautifully designed copy became available before I shared. Please check out the scoping review I co-authored with Eva Susanne Blazek, PhD and Christopher Symons titled "How are Machine Learning and Artificial Intelligence Used in Digital Behavior Change Interventions?" published in Mayo Clinic Proceedings: Digital Health. We term the digital behavior change interventions, or DBCIs, that incorporate AI "AI-DBCIs," and I think what we found might surprise you. We hear so many people talking about the use of #artificialintelligence in apps or other digital health tools in order to coach or change real-world health behavior (i.e., off the screen), but we wanted a more accurate accounting of what's actually happening. From about 3000 papers that met our initial search terms, our inclusion and exclusion criteria winnowed down to just 32 papers covering 23 different AI-DBCIs, the majority of which were developed by academic teams for research purposes. Lirio's Precision Nudging is one of just six commercially available AI-DBCIs with peer-reviewed research available. What did we learn? A few highlights: -Almost every AI-DBCI uses homegrown technology rather than licensing 3rd party platforms. This speaks to the importance of IP and competitive differentiation, but also may get in the way of solid learning about the effects of AI to change behavior, especially given our next finding. -The way behavioral researchers talk about AI varies widely and is often inconsistent with how machine learning experts characterize the technology. Some papers claimed their interventions used AI, but a careful read by an AI scientist suggested a mismatch with modern accepted definitions (for example, hard-coded algorithms that do not learn, or Wizard-of-Oz implementations). Don't take someone's word for it that they are using AI! Rather, we have an opportunity to get educated about AI and very importantly, adopt an accurate shared vocabulary that enhances reproducibility of findings and lets us gather evidence about which AI tools are most fit for purpose across a range of behavioral interventions. -The early evidence on the use of AI to change behavior is promising, but there is SO MUCH work to be done. We need to look at more long-term behavior changes over time, and more rigorous study designs are needed to parse out the role of AI in any behavioral outcomes. There are also so many more behavioral domains where AI could support interventions. It's a wonderful opportunity space. There are more great nuggets in the paper, and we included a plain language glossary of common AI terms and concepts to help our readers become more conversant in discussing AI. Personally I am just excited to have a grounded sense of what exists in terms of AI-DBCIs for real-world behavior and a better roadmap to guide the next months of work in proving out the role of AI in behavioral interventions. Link in comments!