Artificial Intelligence Innovations in Suicide Prevention

Feifan Liu, PhD, FAMIA
UMass Chan Medical School

Presentation: Generalizable and Trustworthy AI for Suicide Attempt Risk Prediction: from Traditional ML to Generative AI

Suicide risk prediction models rarely extend beyond the health systems in which they were developed, creating a critical barrier to cross-site deployment and comparison. This presentation reports findings from the external validation of a state-of-the-art suicide attempt risk prediction model and examines how well locally retrained models generalize across diverse clinical settings. It traces the evolution of methods for analyzing tabular clinical data, from traditional machine learning and deep learning to generative AI, and outlines future directions that shift the focus from predictive performance to measurable improvements in care delivery and expand the scope from structured EHR data to multimodal clinical analysis enabled by agentic AI.

Speaker Bio

Feifan Liu, PhD, FAMIA, is an associate professor of the Department of Population and Quantitative Health Sciences at UMass Chan Medical School and the founding director of the innovative AI for Health (iAI4Health) lab. He is a faculty member of NIH funded P50 CAPES suicide prevention center, co-leading an exploratory project, ADAPT, focusing on suicide risk prediction generalizability and transferability. Trained in computer science, he has expertise in natural language processing, machine (deep) learning, generative AI, and trustworthy AI, with a focus on risk prediction and clinical decision support. Dr. Liu was awarded as an NIH AIM-AHEAD leadership fellow in 2022 and was inducted into Fellow of American Medical Informatics Association (FAMIA) class of 2025. Liu has also served as an AI/ML research mentor for the NIH All of US Research Scholar program and several NIH AIM-AHEAD training programs. He currently serves as an associate editor of npj Mental Health Research, and as a standing member of NIH study section in Clinical AI and Informatics. 

Danielle Mowery, PhD, MS, FAMIA, FACMI
University of Pennsylvania

Presentation: Automated Safety Plan Scoring in Outpatient Mental Health Settings: An Exploratory Study Using Large Language Models

Safety planning intervention (SPI) for suicide prevention intervention involves creation of a written plan to help patients reduce suicide risk. More complete, personalized, and specific SPIs are more effective at reducing suicide risk. However, measuring SPI quality can be labor-intensive and subjective, resulting in clinicians rarely receiving specific, actionable feedback on their use of the SPI. In this presentation, we describe efforts to develop a Safety Plan Fidelity Rater to assess the quality of written safety plans leveraging generative AI.

Speaker Bio

Danielle Mowery, PhD, MS, FAMIA, FACMI, is an assistant professor of informatics in the Department of Biomedical Informatics at the University of Pennsylvania and serves as co-director of the Methods Core for the Penn Innovation in Suicide Prevention Implementation Research (INSPIRE) Center. She is a collaborative investigator that develops natural language processing (NLP), generative artificial intelligence (AI) solutions, data science, machine learning, and computational methods to integrate and analyze information from unstructured texts and structured clinical data to help clinical investigators better understand disease burden, treatment efficacy, and clinical outcomes. Her work aims to uncover scientific discoveries, identify actionable healthcare knowledge, and optimize translation of research into patient care. She is the inaugural Chief Research Information Officer for Penn Medicine—a key position designed to bridge the gaps between clinical data, research expertise, and actionable healthcare knowledge. Dr. Mowery represents Penn Medicine in local, regional, national, and international clinical research informatics communities, i.e., currently, serving as a member of the Association of American Medical Colleges (AAMC) Group on Data and Technology. She is also a Fellow of the American Medical Informatics Association (FAMIA) and Fellow of the American College of Medical Informatics (FACMI). Mowery has extensive training in the sciences that supports her interests in clinical and translational research including a BS in biological sciences, MS in health & rehabilitation sciences, MS and PhD in biomedical informatics from the University of Pittsburgh. She also completed her post-doctoral training at the University of Utah. 

Chris Kennedy, PhD, MPA
Harvard Medical School

Presentation: Pragmatic Machine Learning for Suicide Risk: Routine Assessments, Net Benefit, and AI-Informed Clinical Judgment

Clinical judgment and statistical models are typically evaluated as alternatives for suicide risk prediction. I will argue that they are synergistic and will describe a pragmatic design in which the clinician judges first and a model incorporates that judgment together with routine electronic health record data. The clinician then reviews the estimate for errors, integrates information unavailable to the model, and determines the final treatment plan. The model's inputs include existing structured suicide screeners and risk assessments, whose items operationalize established theories of suicide risk and prevention, so the resulting features measure hypothesized stages and modifiers of a mechanism-informed progression to suicidal behavior. The model's role is to combine those measurements consistently and fairly. At Mass General Brigham, for example, machine learning models built from risk assessments predicted later suicide attempts substantially more accurately than the clinicians' own summary ratings of the same items (Bentley et al. 2025). But accuracy is not clinical utility: decision-analytic net benefit quantifies the extent to which a model would improve clinical decisions across a region of risk where the treatment plan is contested. A prediction improves care only when it motivates a change from the plan the clinician would otherwise have chosen, so its value is concentrated in the cases where the model and the clinician disagree. Both the initial judgment and the final decision depend on the clinician's skill, yet routine practice offers little opportunity to calibrate risk judgments, and formal training in suicide risk prediction or in integrating model estimates into decisions remains uncommon. Providing accurate initial risk judgments and then acting on a divergent model estimate are therefore clinical competencies that benefit from deliberate practice through targeted training and feedback.

Speaker Bio

Chris Kennedy, PhD, MPA, is trained as a biostatistician, data scientist, and psychometrician. His research applies clinical machine learning, deep learning (NLP, computer vision), semiparametric causal inference (targeted learning), and measurement methods to psychiatric problems, particularly EHR-based translational work in suicide, PTSD, emergency utilization, depression, and psychosis.

Quick Info

When: Sept. 30, 2026, 3–4:30 p.m. EDT
Where: Virtual, Zoom
Registration: Registration via Zoom