ABSTRACT. Decades of research to improve the prevention and early detection of suicide risk has largely resulted in the detection of who is most likely to consider suicide, but not when or if that is most likely to happen. Most detection methods presume patients are in contact with the healthcare system, which only reaches a proportion of the at-risk population. Many people at high risk for suicide do not seek professional help because of lack of time, stigma, and fear regarding how they will be treated in the health care system. It is imperative that we develop methods that can identify proximal risk for suicide that does not depend on system- level contact. Web-based search tools are ubiquitous, with 46% of the global population using the internet for information searches and 1.2 trillion searches per year worldwide. Based on our preliminary data, we propose that this online search-engine behavior may prove to be an effective, private, and immediate method of proximal risk detection of suicide for anyone, regardless of their contact with systems of care. We will recruit 1,000 people with mental illness with varying risk for suicide. Participants will provide us access to Google Take-Out (GTO) data, which includes search-engine history and behavior including YouTube. Participants will include those who have report a suicide attempt in the past year (N=500), those who have made an attempt over a year ago (N=250), and those who have thoughts of suicide but never attempted (N=250). All will participate using gold-standard suicide behavior research instruments. Using a case-crossover design, we will evaluate the intermittent exposures (search based proximal risk factors) with an immediate and transient effect on risk and an abrupt outcome (suicide attempt). The case-crossover design is a well-tested and proven approach especially in cases where transient events can trigger acute events such as cardiovascular events, injuries, and death due to environmental exposures and has been studied with interview data to determine warning signs for suicide attempts. Further for predicting suicidal attempt/s, we will use robust ensemble-based machine learning methods such as random forest, gradient boosting to evaluate the predictive nature of qualitative and quantitative features. The study will conclude in a collaborative dissemination planning process with our community partners. Thus, this retrospective and prospective study that aligns GTO data with carefully assessed suicidal thoughts and behaviors has the potential to identify warning signs in search and YouTube data that predict when suicidal risk and lay the groundwork for innovative pathways to suicide prevention.