Topic Modelling Approaches for Identification of Topics within Clinical Notes of Emergency Department Patients with Opioid Misuse
Session Number
Project ID: MEDH 36
Advisor(s)
Dr. Neeraj Chhabra, University of Illinois Chicago
Discipline
Medical and Health Sciences
Start Date
17-4-2024 10:00 AM
End Date
17-4-2024 10:15 AM
Abstract
Opioid misuse is a significant public health challenge in the US, with escalating impacts on emergency medical services and emergency departments. Patients with opioid misuse are often treated as a homogenous population when there are likely subgroups that may influence optimal clinical care. This study aims to investigate and identify these latent groups among patients with opioid misuse. A sample of 1200 UI Health emergency department encounters were retrospectively reviewed and annotated for the presence of opioid misuse using previously published methodology. Of these, 570 cases were positive for opioid misuse. A latent Dirichlet allocation model was then trained on the clinical notes from these patient encounters. Coherence scores for models encompassing 2-19 topics were calculated. The final model was chosen by balancing coherence scores with model complexity. The optimal model was determined to be a 9-topic model with the following topics represented: overdose/altered mental status, skin and soft tissue infection, cardiac disease, limb pain, mental health, critical illness, physical rehabilitation, respiratory conditions, and gastrointestinal/liver disease. These findings highlight the heterogeneity that exists within the population of patients with opioid misuse utilizing the emergency department and suggest that personalized treatment approaches should be investigated to improve patient outcomes.
Topic Modelling Approaches for Identification of Topics within Clinical Notes of Emergency Department Patients with Opioid Misuse
Opioid misuse is a significant public health challenge in the US, with escalating impacts on emergency medical services and emergency departments. Patients with opioid misuse are often treated as a homogenous population when there are likely subgroups that may influence optimal clinical care. This study aims to investigate and identify these latent groups among patients with opioid misuse. A sample of 1200 UI Health emergency department encounters were retrospectively reviewed and annotated for the presence of opioid misuse using previously published methodology. Of these, 570 cases were positive for opioid misuse. A latent Dirichlet allocation model was then trained on the clinical notes from these patient encounters. Coherence scores for models encompassing 2-19 topics were calculated. The final model was chosen by balancing coherence scores with model complexity. The optimal model was determined to be a 9-topic model with the following topics represented: overdose/altered mental status, skin and soft tissue infection, cardiac disease, limb pain, mental health, critical illness, physical rehabilitation, respiratory conditions, and gastrointestinal/liver disease. These findings highlight the heterogeneity that exists within the population of patients with opioid misuse utilizing the emergency department and suggest that personalized treatment approaches should be investigated to improve patient outcomes.