Uncertainty-Informed Decision Making for Biobehavioral JITAIs

April 18, 2024 Brainstorm Outline # Uncertainty-Informed Decision Making for Biobehavioral JITAIs ## Goals for the Brainstorm My goal is for us to outline promising research projects that I will pursue in the near term. I will be submitting a [NIH K99/R00](https://www.nibib.nih.gov/training-careers/training-opportunities/nih-pathway-independence-award-parent-k99r00) proposal by the end of May (deadline June 12). This proposal will outline my primary postdoctoral research aims over the span of two years (K99). This portion is critical to think through because I want to start work on these projects as soon as my current work on simulation test beds is complete. The proposal also outlines continuation work for three years afterwards as an assistant professor (R00). Your help is needed in evaluating / adapting / generating research ideas! ## Questions for you - How compelling do you find each research avenue? What interests you? What would you change about the project to make you more interested? - How do the projects align with our combined expertise? What additional expertise may be needed to conduct the research well? - How practically feasible do you think each project is? How would you improve the execution plan (e.g., methods, data, and/or collaboration team)? - Based on answers above, how would you down-select the projects for a well-scoped set of specific aims? How would you divide the work between K99 and R00 phases? ## Introduction ### Why Make Decisions Informed by Uncertainty? There are pros and cons to maybe everything in life - including JITAIs that tailor their interventions to biobehavioral data. In theory, a JITAI that tailors its interventions based on what it *estimates* of a patient's body or behavior sounds like it could only benefit - the JITAI basically reads your mind and tells you exactly what you need to improve your health in that moment! *But what happens when the JITAI's estimate is wrong?* If a JITAI tells you, "Take a deep breath. You seem stressed," but you are completely relaxed, you can end up losing confidence in the JITAI, or even worse, disengaging. This was a lesson learned from studies like [Sense2Stop](https://doi.org/10.1016/j.cct.2021.106534), where, for the first time, estimates of physiological stress were used to inform digital health interventions. It became increasingly clear that these estimates were not as accurate as we thought they would be - but **we had no way of quantitatively telling when we were more certain about one estimate of stress than another.** And even if we were provided with measures of uncertainty, do we have the decision-making algorithms ready to leverage that information for improved decision-making? ### How can Biobehavioral Insights Help? The term [biobehavioral](https://www.merriam-webster.com/dictionary/biobehavioral) has been [more recently used](https://doi.org/10.1145/3161587.3161591) in the context of biological methods applied to the study of behavior, or behavioral methods applied to the study of biology. The key insight in health contexts is recognizing that *your behavioral and biological health are intertwined*. A related term that is more often used outside of the computing community is [psychophysiological](https://en.wikipedia.org/wiki/Psychophysiology). In many of our contexts (e.g., substance use, physical activity, oral health), biobehavioral insights are leveraged when designing sensors. For example, oral health is typically maintained if one brushes well, so instead of sticking a biosensor in people's mouths to assess oral health, smart brushes assess brush quality as a behavioral proxy. It can be hard to behaviorally estimate how badly someone is experiencing acute opioid withdrawal, but physiologically, it can be easier to estimate increased sympathetic arousal or decreased parasympathetic activity (i.e., "physiological stress"). **I am interested in leveraging these domain science insights to improve our design of estimation and decision-making algorithms.** ## Potential Aims and Projects ### Aim 1: Physiology-informed estimation of state and uncertainty - (K99)"Directionality-informed machine learning" (analogous to [physics-informed machine learning](https://doi.org/10.1038/s42254-021-00314-5)) for model fitting informed by directed acyclic graphs with known relationship signs (i.e., + vs. -) - (R00 Y1-Y2) Improving covariance estimation / uncertainty quantification for downstream estimators via [biosignal quality indices](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5597264/) ### Aim 2: Uncertainty-informed decision-making with state estimates - (K99) Advancing the [measurement error bandit (MEB)](https://arxiv.org/abs/2307.13916) toward a deployable algorithm in practice: uncertainty-informed bandits with predicted context - (R00 Y2-Y3) Either... - Incorporating uncertainty information into JITAI nudges, rather than to decide between nudge or not (R00 Y2-Y3) - maybe here I compare NLP method from possible Aim 3 vs. no uncertainty or reporting uncertainty directly to patient? - Or maybe here I start exploring [decision-focused learning](https://arxiv.org/abs/2307.13565) because I have an optimization problem (bandit) and I have predictors (Aim 1) above... ### Aim 3: Psychology-informed generation of JITAI nudges - (K99 Y2 - R00 Y2) Automating JITAI Nudge Creation Using Calibrated LLMs The idea here is we could potentially use similar methods to calibrating simulation testbed: both are generative models that need to match constraints somewhat determined by domain scientists...But is NLP too far outside of our wheelhouse?