Enhancing Brain Injury Surveillance Using Large Language Models and Topic Modeling

Sun, 19 Oct 2025 00:00:00 +0000

Our presentation, Enhancing Brain Injury Surveillance Using Large Language Models and Topic Modeling, was delivered by Dr. Shockley McCarthy at the AI for BI Conference at The Ohio State University on October 17, 2025.

This work introduces an AI-driven framework for improving the detection and contextual understanding of brain injury cases using unstructured medical and administrative text. By applying large language models (LLMs) and structured topic modeling, we extract interpretable patterns from injury narratives and ICD-coded data to identify mechanisms of injury, intent, and underreported cases of intimate partner violence (IPV)–related brain injury.

Integrating methods from public health, law, and computational linguistics, our approach aims to enhance the precision of injury surveillance systems and inform equitable prevention and policy strategies addressing violence-related traumatic brain injury (TBI).

Download the presentation here .

Detecting Intimate Partner Violence Injury Mechanisms Using LLMs and LDA

Wed, 29 Apr 2026 15:45:00 -0500

Dr. Gia Barboza-Salerno, Olivia McLucas, and Dr. Karla Shockley will present their research at the 2026 Symposium on Data Science & Statistics (SDSS) in Milwaukee, Wisconsin.

Presentation Details

Title: Automated Detection of Intimate Partner Violence Injury Mechanisms from Emergency Department Narratives Using Large Language Models and Latent Dirichlet Allocation

Session: AI and LLM Applications I
Date: Tuesday, April 29, 2026
Time: 3:45 - 5:15 PM CT
Format: 25-minute refereed presentation
Location: Milwaukee, Wisconsin

Registration: Register for SDSS 2026
Registration Deadline: March 25, 2026

Abstract

We developed an AI pipeline designed to overcome limitations in existing surveillance systems by distinguishing the underlying mechanisms of Traumatic Brain Injury (TBI) in intimate partner violence (IPV). The primary goal was to apply a Large Language Model (LLM) to extract and classify behavioral and contextual information from unstructured free-text narratives within the National Electronic Injury Surveillance System (NEISS).

The pipeline first cleans the narratives and uses the LLM to convert the text into six targeted structured semantic representations (e.g., action, cause, diagnosis), which are then quantified using Latent Dirichlet Allocation (LDA) topic modeling. Our results can be used to enhance injury surveillance to identify IPV in emergency department settings.

Introduction

Head, face, and neck injuries are common IPV-related injuries and frequently involve TBI, yet NEISS diagnostic codes lack the behavioral detail needed to differentiate mechanisms of harm. Although NEISS narratives contain brief descriptions of arguments, aggression, object use, and environmental context, they have been underutilized in surveillance research and rarely analyzed at scale.

As a result, diagnostic codes collapse distinct injury pathways—including accidental falls, object impacts, impulsive actions during conflict, and direct assaults—into the same categories, obscuring meaningful variation in risk and limiting clinical screening and prevention efforts.

Our method fills this gap by applying AI-based extraction and topic modeling to detect behavioral and situational information embedded in NEISS narratives, producing a more nuanced understanding of IPV-related injury mechanisms.

Methods

Analytical Framework

We developed an AI pipeline to extract and classify behavioral and contextual information from unstructured NEISS narratives related to IPV:

Text Preprocessing: Used regular expressions to flag domestic-violence–related incidents, expand abbreviations, clean shorthand, and isolate IPV cases
Custom Prompting: Generated prompts by merging cleaned narratives with structured NEISS variables
LLM Extraction: Transformed narratives into six semantic fields (action, cause, diagnosis, mechanism, object, location)
Topic Modeling: Applied LDA to identify latent themes within each semantic field
Typology Classification: Integrated regex-based typology and LDA clusters to produce quantitative injury pathways

Text Processing and Prompting

We used vectorized regular expressions to:

Flag violence-related narratives
Expand abbreviations and clean shorthand
Isolate IPV cases by identifying relationship and conflict terms
Exclude clearly non-IPV incidents

After preprocessing, we generated custom prompts by merging each cleaned narrative with structured NEISS variables to produce context-rich descriptions of the event.

LLM Extraction and Topic Modeling

The LLM transformed cleaned narratives into six structured semantic representations:

Action: What happened (e.g., verbal disputes, punching, striking objects)
Cause: Why it happened (e.g., intoxication, falls, collisions)
Diagnosis: Medical outcome (e.g., fractures, contusions, lacerations)
How: Mechanism of injury (e.g., physical struggle, impulsive acts)
What: Objects involved (e.g., walls, windows, doors, fists)
Where: Location (e.g., home, public place, bathroom)

LDA topic modeling systematically discovered and quantified overarching themes across all responses, resulting in quantifiable dimensions like interpersonal conflict or blunt impact injuries.

Data & Results

Quantification of Contextual Dimensions (LDA Topics)

Topic modeling produced interpretable themes across all semantic dimensions:

Action: verbal disputes, play fighting, punching, striking objects
Cause: intoxication, falls, collisions with walls, windows, or furniture
Diagnosis: fractures, contusions, lacerations, closed head injury
How: physical struggle, impulsive acts, heated altercations
What: walls, windows, doors, bricks, glass, fists
Where: home, public place, outside environment, bathrooms

Identified Head/Neck Injury Typologies

Narrative- and topic-derived clusters aligned with six major typologies:

Typology	%	Representative Narrative
Accidental fall or slip	38.6%	“Slipped and fell hitting head on carpet.”
Self-inflicted harm	19.9%	“Banging head against wall during argument.”
Physical assault	25.1%	“Stabbed in neck with kitchen knife.”
Blunt object impact	9.9%	“Head-butted wall; forehead hematoma.”
Environmental	2.3%	“Pepper spray during fight.”
Vehicle-related impact	1.7%	“Swerved and crashed; head/neck pain.”

Word Cloud Visualization

The figure above shows word-cloud representations of dominant latent topics extracted from LLM-generated semantic dimensions (Cause, Diagnosis, How). Each cloud visualizes high-frequency terms within a topic cluster, highlighting recurrent mechanisms of injury across IPV-related emergency department visits.

Key patterns identified:

Heavy concentration of head trauma in incidents involving walls, floors, or doors
High rates of head contact in intoxication-related falls
Frequent head or neck injury during impulsive self-directed actions

Discussion & Conclusions

Combining regex-based classification with LLM extraction and LDA provides a detailed mapping of how, where, and why brain injuries occur during IPV-related incidents.

Key Findings

Blurred Boundaries: Many injuries categorized as “accidental” or “environmental” occur in the context of relationship conflict, revealing blurred boundaries between unintentional and interpersonal harm
Enhanced Surveillance: These methods improve surveillance and detection of IPV-related injuries in emergency departments by recovering behavioral information that structured diagnostic codes cannot capture
Mechanism Identification: The narrative-based approach identifies specific mechanisms that contribute to injury, whereas structured codes often make it difficult to distinguish between distinct behavioral pathways

Future Work

Expanding prompts to capture richer contextual cues
Validating typologies with larger datasets
Examining variation across sociodemographic and clinical subgroups

About SDSS

The Symposium on Data Science & Statistics is a premier conference bringing together researchers, practitioners, and students working at the intersection of statistics, data science, and computational methods. The 2026 symposium takes place April 28 - May 1 in Milwaukee, Wisconsin.

Keywords: Traumatic brain injury, intimate partner violence, surveillance, Natural Language Processing, topic modeling, emergency medicine, LLM, LDA

Natural Language Processing | ISSUES | The Ohio State University