Project Summary Clinician-delivered relapse prevention interventions for alcohol use disorder (AUD) are effective when delivered but the vast majority of adults with an active AUD do not receive them due to well-known barriers to clinician-delivered treatment. Digital therapeutics (smartphone “apps” that are used to prevent, treat, or manage a medical or psychiatric disorder) can address these barriers. Unfortunately, the benefits from digital therapeutics may be constrained because engagement with them is often not sustained or matched to patients’ needs. The next wave of “smart” digital therapeutics that include embedded machine learning lapse prediction models powered by personal sensing can address these constraints by guiding patients to sustain engagement with the specific interventions and supports that are most personally risk-relevant and therefore most effective. Personal sensing has been possible within digital therapeutics for AUD for several or more years. Machine learning lapse prediction models are emerging now, and the models developed by our team meet or exceed performance thresholds necessary for useful clinical applications. We are well-positioned to develop a smart version of our Center’s A-CHESS digital therapeutic by embedding our lapse prediction model into an existing version of our digital therapeutic that already has sensing capabilities. However, we must first determine how best to provide model feedback to patients so that they use this information and follow its recommendations. In this application, we propose to optimize feedback from our lapse prediction model (via daily engagement messages) both to increase risk-relevant engagement with Smart A-CHESS and to improve clinical outcomes over six months among 416 participants with moderate to severe AUD. Following the Multiphase Optimization Strategy, we factorially manipulate four candidate components of these daily engagement messages that convey transparent, individualized, risk-relevant information from our machine learning lapse prediction model to participants. These message components include: 1) lapse probability, 2) lapse probability change, 3) important model features, and 4) a risk-relevant module recommendation. These components use output that would be available from any machine learning lapse prediction model such that conclusions about the impact of these components on engagement can generalize beyond our specific machine learning model. Similarly, engagement messages including these message components could be used in any smart digital therapeutic for AUD, allowing conclusions to generalize to current and future variants of smart digital therapeutics for AUD. At the conclusion of the grant period, we will also deliver this optimized smart digital therapeutic as a tangible product and model for how to embed sensing and machine learning into other existing digital therapeutics.