Leveraging Large Language Models and Machine Learning Algorithms to Assess Depression and Anxiety Symptoms and Risks for Patients with Cardiovascular Disease or Diabetes Mellitus

Project Summary. Depression and anxiety are 2-4 times as likely prevalent among cardiovascular disease (CVD) or diabetes mellitus (DM) patients than among those without CVD or DM. Co-morbid depression and anxiety have a detrimental impact on CVD or DM patients, including exacerbating chronic symptoms and

increasing mortality. However, co-morbid depression and anxiety are often underdiagnosed due to the multi- layer barriers at the patient, clinician, and health system levels. Particularly, symptomatic issues and care needs for depression and anxiety might not be easily shared during cardiology or endocrinology visits while

clinicians focus on chronic physiological symptoms. The patient portal allows patients to communicate with providers to share their symptoms and concerns, which may signal the early signs of depression and anxiety. Recently introduced Large Language Model (LLM) algorithms have created a robust environment for extracting

meaningful topics from large text data. Moreover, machine learning (ML)-based risk models have been designed to predict the risk of CVD or DM, yet, modeling to predict the risk of co-morbid depression and anxiety has been remarkably rare. Thus, in Aim 1, Dr. Kim will identify symptomatic issues and care needs for

depression and anxiety among CVD or DM patients using patient portal messages. More than 46 million messages from Stanford Health Center (SHC) will be analyzed by LLM algorithms. It will transform the raw text data into groups of words, then weight them to generate salient topics which represent the primary symptoms

and care needs. The generative AI algorithm will enhance interpretability of the topics. In Aim 2, Dr. Kim will develop co-morbid depression and anxiety risk prediction models and specify risk factors among CVD or DM patients. She will leverage the Least Absolute Shrinkage and Selection Operator algorithm, using the electronic

health records of more than half a million patients at SHC to calculate the area under the curve to present the accuracy of prediction and odds ratios with 95% confidence intervals to indicate the strength of risk factors. The long-term goal is to apply this patient portal-based symptom detection and risk prediction approach to

other at-risk populations to prepare tailored interventions to ultimately improve depression and anxiety outcomes, aligning with the mission of NIMH, "to transform the understanding and treatment of mental illnesses, paving the way for prevention, recovery, and cure." The Career Development Plan will enable Dr.

Kim to gain hands-on skillsets to use the newest LLM packages and construct LASSO-based prediction models independently, with an advanced understanding of the clinical context of mental disorders under the guidance of mentors (Dr. Linos in Digital Health, Dr. Rodriguez in Psychiatry, Dr. Hernandez-Boussard in

Medical Informatics) and advisors in Biostatistics, Cardiology, Endocrinology, Bioethics. All in all, the strong mentor team and solid training plans along with an excellent institutional support, will fully prepare Dr. Kim to be a well-disciplined independent investigator in computational epidemiology and mental health.