Suicide, depression, and anxiety screening using an automated virtual assistant - Summary: Each year in the US, more than 47,000 people die by suicide and many of these tragedies could be prevented with early identification and treatment. At-risk individuals rarely receive screening until depression, anxiety, and suicide symptoms are severe. Universal screening tools deployed in schools and physicians’ offices could address this problem, but the lack of an objective tool along with a shortage of mental health clinicians in these settings is a major barrier to screening on a public health scale. Current methods are subject to inter-rater variability, require experience for interpretation, and are somewhat cumbersome to administer. Meta-analyses have indicated that predicting suicide death using current methods does no better than random chance. There is a need for a singular, easy-to-use, accurate, reliable, objective clinical decision support tool that can be used as a clinical and public health approach to identify patients at risk of suicide, depression, and anxiety to guide at-risk individuals to life-saving resources and interventions. The complexity and broad variability of so- called “thought markers” such as verbal and non-verbal language cues resist easy classification by a clinician, but artificial intelligence (AI) approaches can identify consistent, meaningful patterns. By using a natural language processing AI approach, we have developed Clairity™, an AI-enabled app that automates the identification of at-risk individuals based on speech and vocal markers. Initial development has demonstrated Clairity’s potential for predicting suicide risk, depression, or anxiety, but it currently relies on a live interviewer. Perpetual health care worker shortages, exacerbated by the pandemic, present a major barrier to scalability. Interactive voice response (IVR) may be a solution to health system strain that has no signs of abatement. An IVR system could also reduce inter-interviewer variability and, for some patients, entice better engagement. We will assess the effects of user perception of IVR as well as demographics on model performance when using an IVR system in comparison to human interviews. We will validate modified models to inform key decisions regarding our commercialization path and future trial designs. Aim 1. Characterize the effects of IVR data collection on model performance. Factors that decrease model performance by ≥ 10% AUC versus live interviewer will be identified for model improvement. Aim 2. Validate updated algorithms with a holdout data set. Successful models will correctly classify risk categories with an AUC ≥ 80%. We will also analyze demographic differences to inform tuning of the IVR system (e.g., longer pause times, allowing participants to select a specific virtual assistant). Aim 3. Assess user perspective on engagement with IVR. Patient perceptions of virtual versus live interviewers may affect model performance. We will survey patients regarding their perceptions and emotions in dealing with a virtual agent and assess correlation with model performance. Impact: Advances from this project are expected to enable a scalable, real-time clinical decision support tool for risk screening that could improve treatment referral and ultimately reduce suicides.
