Co-Pilot for Health: Personalized Algorithmic AI Nudging to Improve Health Outcomes
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: We're looking at the paper: "Co-Pilot for Health".
Title and authors: Jane: So we’re diving into the paper "Co-Pilot for Health: Personalized Algorithmic AI Nudging to Improve Health Outcomes," which is led by Jodi Chiam, Aloysius Lim, Cheryl Nott, Nicholas Mark, Sunil Shinde, and Ankur Teredesai. We talked about the general idea of personalized nudging earlier.
Tom: Right. I want to focus on how this specific setup—using a GNN RecSys on wearable data—actually translates into better health actions for the people they tested.
Lu: The system ingests participant data daily to generate nudges that are contextually relevant, meaning the system learns relationships between users, nudges, and their attributes through a knowledge graph constructor.
Meng: Learning those relationships dynamically sounds smart, but how much data is this actually talking about? Are we talking about millions of people or just a pilot group?
Lalam: The paper describes NudgeRankTM which uses a Graph Neural Network to learn relationships between users, nudges, and their attributes by ingesting participant data from various sources daily to generate personalized nudges that are contextually relevant.
Jane: It sounds like they’re tying in with existing health infrastructure too, which is interesting because it grounds the research in real-world usage.
The paper's summary: Tom: The main result they found was that these personalized nudges led to statistically significant improvements in physical activity for both groups compared to control groups over the entire twelve weeks <ref:2401.10816#pg3>.
Jane: What did those improvements look like in concrete numbers? Did people just feel slightly better, or were there measurable changes?
Tom: Over those twelve weeks, participants who got nudges showed a two point two percent more daily steps compared to just three thousand eight hundred twenty-nine versus three thousand seven hundred forty-seven steps at p equals one point six eight times ten to the power of negative five. And they also had a two point seven percent more minutes of moderate to vigorous physical activity per week with a p value of nine point zero seven times ten to the power of negative three.
Lu: Those numbers show that this effect was larger for Group one participants, who were only in the National Steps Challenge program, showing a six point one seven percent more daily steps with a p value of three point zero nine times ten to the power of negative four <ref:2401.10816#pg2>.
Meng: So it worked even for people who were already relatively active, not just beginners? That's important because if it doesn't work for everyone at scale, we can't deploy it widely.
Lalam: The system dynamically adapts by tracking daily fluctuations in behavior to ensure the daily nudges remain contextually relevant for the participant.
Jane: And they also mentioned how well people actually engaged with these messages themselves? That’s a crucial part when you’re talking about digital intervention.
The paper's improvements: Tom: The engagement metrics were pretty solid too; the overall open rate was thirteen point one percent, and they noted that nudges opened were rated useful six times more often than not useful by participants <ref:2401.10816#pg3>.
Lu: They also showed that the system adapts by tracking daily fluctuations in behavior to ensure the daily nudges remain contextually relevant for the participant. That’s a key part of making sure it doesn't just spam people with irrelevant stuff.
Meng: I’m wondering about the deployment cost and how this works at scale, because if you have millions of users, running this GNN on every single daily data point sounds computationally expensive.
Tom: The system architecture uses a Kubernetes cluster for distributed computing where the Knowledge Graph Constructor processes all ingested participant and nudge data into a heterogeneous knowledge graph that captures the dynamic interactions between participants, nudges and other knowledge about them.
Lu: Then the NudgeRanker generates the ranked list of nudges for each participant using a Knowledge-Graph powered RecSys that employs a Graph Neural Network to learn relationships. After that comes the Constraints Filter module which applies several filters, such as nudge budgets and removal of nudges that participants had rated “not useful.”
Jane: So they’ve got a whole pipeline from data ingestion through ranking, filtering for relevance, and finally personalizing the content before it hits the user’s screen.
Conclusion: Tom: To wrap up this discussion on "Co-Pilot for Health: Personalized Algorithmic AI Nudging to Improve Health Outcomes," the paper shows that personalized nudges using a Graph Neural Network RecSys can generate algorithmically targeted nudges to encourage healthy behaviors amongst participants.
Lu: The key finding is that this approach, by leveraging granular health behavior data from wearable fitness devices, allows the system to shape health behaviors of large populations automatically across different wearable types and disease conditions at scale.
Meng: But they did have a limitation, which I want to highlight—the paper notes that participants had to open the mobile app on their phones to upload tracker data, which resulted in partial censoring of activity data.
Lalam: They also pointed out that they need longer studies to understand whether digital nudging leads to long-term habituation of behaviors.
Jane: It seems like a really promising start, showing a clear path for using AI and personalized recommendations to nudge people toward better health habits based on their actual activity levels.
Tom: Exactly. The potential here is shaping health behaviors at scale, which has tremendous potential to improve global health outcomes, but those long-term habituation questions still need answers.
Jodi Chiam, Aloysius Lim, Cheryl Nott, Nicholas Mark, Sunil Shinde, Ankur Teredesai
CueZen, Inc. · Department of Critical Care Medicine, Swedish Medical Center, University of Washington
cs.HC, cs.AI, cs.LG
Submitted: 2024-01-19
Updated: 2026-11-18
Code: https://github.com/xiangwang1223/knowledge
Importance score: 91/100
The gist: The paper describes an AI-driven platform for digital algorithmic nudging enabled by a Graph-Neural Network based Recommendation System and granular health behavior data from wearable fitness devices…
Key concepts
- NudgeRank™
- This is the novel recommendation system built using a Graph Neural Network (GNN). It learns complex relationships between users, available health nudges, and participant attributes to generate highly personalized recommendations tailored to each individual's context.
- Graph Neural Network (GNN)
- A type of deep learning model used here that processes data structured as a graph. In this system, the graph connects participants, nudges, and their attributes, allowing the AI to understand how different elements are related and suggest relevant actions.
- Knowledge Graph Constructor
- This module ingests all participant and nudge data daily to build a dynamic knowledge graph. This structure captures all interactions between users and nudges, providing the rich context needed for the recommendation system to make informed decisions.
Terminology
Summary
The paper describes an AI-driven platform for digital algorithmic nudging enabled by a Graph-Neural Network based Recommendation System and granular health behavior data from wearable fitness devices to improve health outcomes at scale The ability to shape health behaviors of large populations automatically, across wearable types and disease conditions at scale has tremendous potential to improve global health outcomes
Study Objective and Methodology
The study aimed to assess the impact of digital personalized nudges on improving health behaviors and participant engagement The researchers designed a novel RecSys called NudgeRankTM which uses a Graph Neural Network (GNN) to learn relationships between users, nudges, and their attributes This system ingests participant data from various sources daily to generate personalized nudges that are contextually relevant The design involved a 12-week public beta where participants received one personalized nudge per day targeting either steps or MVPA Participants were randomly selected from the existing pool of active participants enrolled in the National Steps ChallengeTM (NSC) program
Efficacy of Personalized Nudges
The personalized nudges led to statistically significant improvements in physical activity for both groups compared to control cohorts Over the entire 12-week period, participants who received nudges showed a 2.2% more daily steps (3,829 vs. p. 3 3,747 steps; p = 1. p. 368×10−5) and a 2.7% more minutes of MVPA per week (85. p. 20 vs. 82.8 minutes; p = 9. p. 007×10−3) than those who did not get personalized nudges These effects were larger for Group 1 participants, who were participants of the NSC program only, showing a 6. p. 317% more daily steps (3,088 vs. p. 3 2,909 steps; p = 3. p. 309 × 10−4) The increase in physical activity observed was persistent across the weeks, with a statistically significant increase in average daily steps for all participants over the 12 weeks of study
Nudge Engagement and Personalization
The engagement metrics demonstrated that nudges were very well received by participants The overall open rate was 13.1%, with 96 nudges authored by HPB subject matter experts in the Nudge Library Among the nudges opened, they were rated useful 6 times more often than not useful, with 17,111 participants rating them useful compared to only 2,790 rating them not useful The system dynamically adapts by tracking daily fluctuations in behavior to ensure the daily nudges remain contextually relevant for the participant
Clinical Significance and Limitations
The positive effects of nudging were observed in both participants who started at lower levels of physical activity (Group 1) and those who were already more active at the beginning (Group 2) The increase in physical activity observed was statistically, and likely clinically significant as the effect size could be expected to have salubrious effects based on prior studies A limitation noted was that participants had to open the mobile app on their phones to upload tracker data, resulting in partial censoring of activity data Longer studies are needed to understand whether digital nudging leads to long-term habituation of behaviors
System Architecture Details
The NudgeRankTM system is composed of several modules operating in tandem, leveraging distributed computing technologies running on a Kubernetes cluster The Knowledge Graph Constructor processes all ingested participant and nudge data into a heterogeneous knowledge graph that captures the dynamic interactions between participants, nudges and other knowledge about them The NudgeRanker generates the ranked list of nudges for each participant using a Knowledge-Graph powered RecSys that employs a Graph Neural Network (GNN) to learn relationships Finally, the Constraints Filter applies several filters, such as nudge budgets and removal of nudges that participants had rated “not useful” The final list of nudges is processed by the Content Personalizer to render any personalized fields in the nudge templates with the participant’s own data
Data Availability and Acknowledgements
All summary data generated or analyzed during this study are included in this published article and its supplementary information files Data of individual participants collected or analyzed during this study are not publicly available due to privacy The Health Promotion Board of Singapore (HPB) provided the Healthy 365 mobile application and data for this study HPB approved the study design and collected the data The authors declare no nonfinancial competing interests The authors designed the underlying algorithms and AI solution, while others designed the study and collected and analyzed the data
References
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Personalized algorithmic nudging using a Graph-Neural Network RecSys can generate
algorithmically targeted nudges to encourage healthy behaviors amongst participants,
which allows for contextually relevant messaging based on real-time behavior changes. -
The AI system can ingest and process dynamic data, including
daily health behaviors that are synced to the mobile application, demographic or other social determinants of health, health history, program enrollment and participation data,
to update a knowledge graph daily. -
The system will compute the
up-to-date list of candidate nudges for each participant
by leveraging the Knowledge Graph and Knowledge-Aware Attention from KGAT to considerhigher-order relationships between different entities in the graph.
-
The system implements dynamic relevance tracking where it adapts nudge sets based on behavior changes, as evidenced by the capability to target participants who
walk less than 5,000 steps per day
and automatically update candidate nudges. -
A Constraints Filter module ensures
nudge budgets (maximum number of nudges per day) are applied to minimize nudge fatigue,
and removes nudges that a participant had ratednot useful,
ensuring high relevance and preventing annoyance.
Related papers
- EmoPhone: A Multi-Wave Dataset for In-the-Wild Mobile and Wearable Affect Sensing
- Tasteply: Externalizing Stable and Situational Preference as Manipulable Spatial Objects
- Shielded Trajectory Sampling for Personalized Relational Safety in Conversational AI
- Reconstruction and Reflection of Positive Experiences through Resurfacing Laughter-indexed Everyday Moments
- LLM-Assisted Reranking to Operationalize Nuanced Objectives in Recommender Systems
- Explainable Iterative Data Visualization Refinement via an LLM Agent