Suicide is a leading cause of death, and individuals who attempt suicide and receive hospital treatment are at high risk for suicide within a year. The identification and validation of warning signs (WS) for suicidal behavior – near-term risk factors– is a national priority. Determining if an individual is at risk now drives high-impact decisions in acute care settings within emergency departments (e.g., whether to admit a patient) and crisis lines (e.g., whether to send a mobile crisis team). Yet, there has been little research on ‘when’ individuals are at near-term risk or WS (i.e., within minutes, hours, a day) for suicide attempts. This clinically- and theoretically-driven study addresses critical gaps in our understanding of WS for suicide attempts. We seek to a) discover novel warning signs candidates for suicide attempts, b) validate, and generate the first risk estimates, for these candidates and WS put forward in recent theoretical formulations, c) compare risk- estimates of WS to determine if those currently prioritized in risk assessments in acute care settings is warranted, and d) develop new algorithms to detect linguistic signals of specific WS content in patients' narrative interviews. We propose a multi-site mixed-methods study that will recruit 400 adults currently hospitalized for a suicide attempt in two academic medical centers in the Upper Midwest. Subjects will be asked to tell the narrative story of their attempt in their own words, and also undergo a detailed semi-structured interview to obtain systematic data about hypothesized WS on the day of the attempt and the day prior. We will discover potential novel WS candidates using subjects’ narrative stories coded by experts using qualitative methodology (Aim 1). Next, we will validate a priori and novel candidate WS (Aim 2). Case-crossover methodology will be used, a within-subjects design that uses subjects as their own control. The semi-structured interview data are analyzed through comparisons of the presence/intensity of hypothesized WS on the day of the attempt (high-risk case period) to the day prior (lower risk control period). Finally, we will develop and test an algorithm to detect linguistic signals of specific WS content (Exploratory Aim 3). Natural language processing and deep learning models of language will be used to detect WS within the narratives. WS for suicide attempts are extraordinarily difficult to study due the practical challenge of examining the hours preceding an act of suicide. The project uses innovative qualitative and quantitative methods to address this challenge in a rigorous fashion. The study is designed to provide scientifically grounded WS to inform clinical decision-making, patient/family education, and automated risk identification.