PAL for Behavior Change

Habit change is the future of health care, self-care, and self-change

Goals

1. Create an evidence-based and context-aware platform for in-the-wild behavior monitoring and behavior change interventions in the real world.
2. Implement reinforcement learning-based closed-loop behavior change interventions for personalized behavior change support.
3. Provide interpretable human-in-the-loop models for low-shot and personalized context detection, context prediction, and behavior change interventions.

Motivation

Healthy behaviors can help us not only prevent and manage health problems, but also achieve self-fulfillment and self-actualization. Behavior change is hard, however, as humans are constrained by their cognitive and emotional constraints. Humans have two types of thinking [Thinking, Fast and Slow]: system I is automatic and effortless, and system II is thoughtful and effortful. Many of our everyday actions are automatic and when automatic actions are repeated in stable contexts, they become habits, i.e., automatic and efficient responses to stable contexts. Successful behavior change, thus, requires context-aware habit change, but traditional behavior change resources, e.g., self-help books and therapists, are not present with us in our everyday lives.

Principles

1. Enable, not enforce: PAL supports a range of behavior change goals and leverages the intrinsic motivation of users to pick their goals, interventions, contextual triggers, and self-tracking sensors. PAL’s self-tracking, goal planning, and contextual reminders aim to assist and empower everyday self-change. 
2. Supplement, not substitute: PAL supports habit formation by anchoring new behaviors in existing routines or by breaking old habits using context-aware reminders. Healthy habits enable long-term and sustainable behavior change so the users train their behavior change muscle and are not dependent on PAL. 
3. Connect, not confine: PAL makes behavior change an inclusive, not an isolated, process. Using PAL, users can exchange audio/text behavior support messages to serve as behavior change reminders and can also share their behavior change progress and goals. PAL aims to be an extension of ourselves and our support system to provide real-time behavior change support.

Platform

PAL’s behavior change platform has three key components: 

1. Wearable Device: The wearable device (Fig. 1) consists of self-tracking sensors (camera, heart-rate, and Inertial Measurement Unit), user input (microphone, button, and tap), user output (open-ear audio), and on-device deep learning accelerator (Google Coral USB Accelerator).  
   

2. Mobile + Web App: The web app has a habit-setting interface (Fig 2) for selecting behavior change goals, reminders and contextual triggers, and a visualization interface (Fig 3) for viewing raw sensor data and machine learning models results. The mobile app tracks user’s location, physical activity, and app usage, and allows the users to customize PAL’s sensors (Fig. 3). 

3. Machine Learning Models: We have machine learning models for Context Detection, Context Prediction, and Intervention Suggestions. The models are modular and hierarchical, and we use knowledge graphs to induce relational biases and prior knowledge into the models. Our models are interpretable and tunable for human-in-the-loop learning. Users can label unrecognized or incorrectly recognized contexts for semi-supervised and active online learning. For context detection, we combine results from mobile-optimized on-device and transfer learning-based low-shot and personalized models for context detection. Fig 4 shows some results of PAL’s on-device models. As we deploy PAL in the real world, we aim to evaluate different reinforcement learning strategies for context-aware and personalized behavior change interventions. 

 Results

We have conducted preliminary in-the-wild full-evaluations of PAL’s wearable device and machine learning models, and a user survey (n=51) of PAL’s behavior change affordances and mobile and web app interface. Our findings show that the users find PAL useful and usable for behavior change (Figure 5 shows some results).