Researchers at Stanford, Harvard, and the University of Michigan propose an AI framework they claim can be used to deliver recommendations via smartphone that encourage healthier lifestyles. By personalizing the recommendations and making data on other users available to bolster training, they claim their approach reduces regret — i.e., the number of actions taken when there was a better choice in hindsight — by 26%.
Mobile health apps aim to support healthy behaviors by offering opportunities to impact health across a range of domains. For example, a health app might send walking suggestions at times and in contexts (e.g. weather, current physical activity, location) when a person is likely to follow suggestions. But the effectiveness of any health app requires providing recommendations at useful times while avoiding overtreatment that might lead to disengagement.
The researchers’ system, dubbed IntelligentPooling, aims to learn an optimal policy for when and how to intervene for each person and context. It indexes decision times — i.e., times when a treatment could be provided — on a per-user basis and has users choose health recommendations throughout the course of a day. Over time, the system develops personalized treatment policies for each user, algorithmically learning from data pooled from the users’ devices.
The researchers conducted a study involving 10 subjects with Fitbit Versa smartwatches as part of a larger trial intended to optimize Stage 1 hypertension interventions. Activity suggestions were randomized five times per day for each participant throughout the 90-day trial, tailored to sensor data like location, weather, time of day, and day of the week. IntelligentPooling determined whether to make an activity suggestion based on 107 data points from the 10 users; autonomously, it could decide to send (or not send) a notification with a suggestion.
The researchers report that overall, IntelligentPooling suggested the “full range” of available treatments despite sending them less frequently compared with a baseline. They note that the study was relatively small and that they can’t claim IntelligentPooling improved health outcomes — that would require larger studies. But they assert the system can overcome some of the challenges faced when learning personalized policies in limited data settings.
“When data on individuals is limited a natural tension exists between personalizing (a choice which can introduce variance) and pooling (a choice which can introduce bias),” the researchers wrote. “We view adaptive pooling as a first step in addressing the trade-offs between personalization and pooling. The question of how to quantify the benefits and risks for individual users is an open direction for future work.”
