Integrating Loneliness and Social Isolation Insights into Late-Life Suicide Risk Prediction Through the Digital Phenotyping of Electronic Health Records

ABSTRACT The crisis of loneliness and social isolation, particularly among the older adults, alarmingly correlates with an increased risk of suicide. Although this crisis is particularly prevalent among older patients, current healthcare systems face challenges in effectively monitoring late-life loneliness, social isolation and associated suicide risks, partly due to insufficient and underdeveloped automated surveillance practices. According to recent studies, increased healthcare utilization is observed among this population prior to suicides, underscoring a critical window for timely intervention and support that remains unexploited, despite available patient data. Our research seeks to bridge this gap by innovatively interrogating electronic health records (EHRs) with rigorous, interpretable artificial intelligence (AI) methods, including those involving large language models (LLMs). Our approach is aimed at significantly enhancing the precision of suicide risk predictions among older adults grappling with isolation and loneliness. Contemporary AI models employed in healthcare settings often overlook the rich narrative data embedded in EHRs, which vividly capture the nuances of patients' experiences with loneliness and social isolation. This is partially due to the fact that until recently natural language processing (NLP) methods lacked the capabilities to detect the often variable and distributed lexical expressions used to describe complex clinical concepts. To address this gap, our project aims to develop advanced prediction models that integrate social factors, specifically loneliness and social isolation, through longitudinal EHR phenotyping. Our project is in line with the National Institute of Mental Health's (NIMH) emphasis on computational methods to detect patterns linked with social isolation and suicidal tendencies in older adults. Utilizing the prowess of LLMs, we will delve into clinical EHR texts to detect indications of loneliness and social isolation in individuals aged 55 and above, conducting differential analyses across different age brackets to gain a deeper understanding of how these factors influence suicide risk in later life. The project comprises three main aims: 1) To develop and validate a novel tool that identifies linguistic markers of loneliness and social isolation in clinical notes; 2) To integrate insights derived from the tool into a suicide risk prediction model, optimized for older populations; and 3) To undertake a comprehensive evaluation/analysis of bias and interpretability of AI models to mitigate the risk of biases in AI algorithms. The outcomes of our project are expected to enhance the use of EHR data for better understanding and prediction of risks linked to loneliness, social isolation, and late-life suicide, and support monitoring efforts through an interactive dashboard. We will release our tools and models as an open-source toolkit, enabling broad application, validation, customization and deployment. The proposed research will strategically address a high-risk, high-reward problem, with high-utility deliverables in each aim. The overarching objective is to foster more responsive and efficient learning healthcare systems, with future research building on the outcomes and exploring external validations and implementation across institutions nationally.