Machine Learning | ISSUES | The Ohio State University

Modeling Firearm Injury Risk Using Spatial Random Forests

Thu, 23 Oct 2025 00:00:00 +0000

Our presentation, Modeling Firearm Injury Risk Using Spatial Random Forests, will be delivered by Dr. Gia Barboza-Salerno at the Women in Statistics and Data Science on November 12, 2025 in Cincinnati, OH.

This study introduces a spatial machine learning framework for predicting firearm injury risk across Chicago census tracts. Using Spatial Random Forest Modeling (SRFM), we integrate social, built, and natural environment indicators, such as social vulnerability, housing burden, green space, traffic volume, and tree canopy, to identify localized structural determinants of firearm violence.

Results demonstrate that racial composition, economic hardship, and environmental stressors jointly predict spatial clustering of firearm injury, while local variable importance reveals that built environment features exert context-specific effects across neighborhoods. These findings highlight the value of spatially aware, data-driven approaches for informing equitable, place-based violence prevention and policy.

Abstract: Gun violence remains a critical public health crisis, particularly in urban centers like Chicago, where firearm-related injuries are among the highest in the nation. These injuries are not randomly distributed; they are spatially concentrated in neighborhoods shaped by structural inequality, including racial segregation, economic disinvestment, and environmental neglect. While random forest models have shown promise in predicting firearm violence and identifying neighborhood-level risk factors, conventional approaches often ignore spatial dependence, leading to biased estimates, inflated variable importance, and reduced geographic generalizability. This study applies spatially explicit machine learning to model non-fatal shootings across Chicago census tracts from 2020 to 2024. A curated set of predictors was compiled, including measures of social and environmental vulnerability, the built environment, and housing inequality. Three research questions guided the analysis: (1) Does incorporating spatial structure via Moran’s Eigenvector Maps (MEMs) in spatial random forests and spatially structured random effects in a Bayesian spatial model improve predictive accuracy compared to non-spatial models? (2) Which neighborhoods exhibit elevated firearm violence risk due to unmeasured spatial processes, as captured by spatial random effects? (3) Does prediction error vary across geographic space, revealing limits to model transferability? Findings suggest that the spatial random forest model outperformed the non-spatial model (pseudo R² = 0.98 vs. test R² = 0.50). Spatial cross-validation revealed geographic variation in accuracy (median nRMSE = 0.76), while spatial random effects from a Bayesian Poisson model identified unexplained latent risk. The enhanced predictive accuracy reveals hidden geographic risk patterns, suggesting that social, structural, and environmental factors can efficiently inform targeted violence prevention strategies.

Download the presentation here .

Disability-related depression inequities are driven by accessibility deficits: Evidence from crowdsourced sidewalk data and machine learning

Tue, 03 Nov 2026 14:45:00 +0000

Session: Mobility, Transit, and Built Environments for Disabled People
Program: Disability
Date: Monday, November 2, 2026 | 2:45 PM – 3:00 PM

Identifying place-specific mortality mechanisms by integrating cause-of-death narratives and neighborhood typologies using machine learning

Mon, 02 Nov 2026 09:15:00 +0000

Session: Epidemiology Section: Conceptualization and Measurement
Program: Epidemiology
Date: Monday, November 2, 2026 | 9:15 AM – 9:30 AM

Predicting harsh school discipline in Albuquerque Public Schools: A machine learning analysis with temporal validation

Mon, 02 Nov 2026 08:30:00 +0000

Session: Public Health Social Work Perspectives on Health and Well-Being Across the Lifespan
Program: Public Health Social Work
Date: Monday, November 2, 2026 | 8:30 AM – 10:00 AM

Detecting Hidden Mechanisms of Abuse-Related Injury in Children: Structured Topic Modeling of Clinical Death Reviews and Legal Opinions

Mon, 24 Aug 2026 00:00:00 +0000

Presented at the ISPCAN World Congress on Child Abuse and Neglect, Melbourne, Australia, August 2426, 2026.

Full title: Detecting Hidden Mechanisms of Abuse-Related Injury in Children: Structured Topic Modeling of Clinical Death Reviews and Legal Opinions Triangulating Multidisciplinary Child Fatality Reviews with Factual and Procedural Narratives from Case Law

Abstract

Existing child maltreatment surveillance systems miss or misclassify between 8090% of maltreatment deaths, and CPS records undercount fatalities by 5576%. Unstructured narrative text clinical death reviews, court opinions, and social work records accounts for 8090% of all organizational data produced, yet it is rarely leveraged for injury surveillance.

This presentation introduces a two-study AI + Structured Topic Modeling (STM) pipeline that integrates two distinct narrative corpora to recover latent, abuse-related injury mechanisms that structured administrative codes routinely obscure.

Study Design

	Study 1: CFRT Death Reviews	Study 2: CourtListener Case Law
Source	Colorado Child Fatality Review Team reports	Appellate opinions retrieved from CourtListener
Documents	223 child-level narrative records	751 opinions → 335 confirmed cases from Ohio
Text	Two free-text fields per report (clinical + administrative)	Facts + procedural history
LLM Extraction	GPT-4o-mini: caregiver behavior, injury mechanism, setting, diagnosis	GPT-4o-mini: 9 structured fields including perpetrator, victim type, legal outcome
STM Models	Risk & Contributing Factors; Circumstances	Injury Mechanism; Long-Term Consequences
Covariate	Incident year (temporal prevalence)	Judicial ruling (affirmed / reversed)

Research Questions

Can large language models transform unstructured child fatality narratives into standardized semantic features caregiver behavior, injury mechanism, setting, diagnosis?
What latent themes emerge, and what typology of risk and incident circumstances do they suggest?
Do the recovered themes reveal recurring patterns of head and neck trauma consistent with abuse-related injury?
How does the prevalence of these themes shift across incident years (20202025)?

Key Findings

Both corpora converge on craniofacial injury as the dominant mechanism routinely undercounted by ICD codes.
Strangulation emerged as a critical lethality predictor; soft-tissue damage can surface 23 days post-injury, requiring clinician and court training on delayed sequelae.
Records document long-term neurological harm (cerebral palsy, pituitary damage, blindness) in children often years before clinical follow-up.
Children represent a structurally distinct injury profile in both datasets, warranting specialized clinical and legal protocols.
“Accidental” injuries (falls, environmental contact) are frequently embedded in an abuse context screening for mechanism, not just diagnosis, is essential.
Appellate opinions and death reviews function as injury-surveillance sources beyond structured administrative fields.

Methodology

The pipeline applies:

PDF parsing (pdftools) and regex isolation of narrative fields
GPT-4o-mini for LLM-based structured field extraction
Structured Topic Modeling (STM) with incident year or judicial ruling as a prevalence covariate
BART zero-shot classification as an independent supervised-style validation check
FREX word scoring (frequency � exclusivity) for topic labeling

Lab Members

This presentation was developed by the ISSUES Lab team at The Ohio State University:

Dr. Gia E. Barboza-Salerno (Founding Director), Dr. Henriikka Weir, Malcolm McCarthy, Dr. Karla Shockley McCarthy, Dr. Amy Watson-Grace, Dr. Keith Warren, Esenia Cassidy, and Olivia McLucas with collaborators Scottye Cash and Suzanne Perkins.

📄 Presentation

Download the presentation slides 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

Living Here Predicts How You Will Die: Neighborhood Context and Mortality Patterns

Wed, 04 Feb 2026 00:00:00 +0000

Abstract

Understanding how neighborhood context shapes patterns of death is critical to identifying the spatial distribution of health inequities. Medical examiner data contain rich narrative descriptions of causes and circumstances of death that are often absent from clinical records and can offer novel insights for public health prevention.

This study integrates Geographic Information Systems (GIS), machine learning (ML), and natural language processing (NLP) frameworks to demonstrate how neighborhood clustering and predictive text analytics can jointly reveal spatially stratified mortality patterns to examine how structural neighborhood conditions predict not only whether but how people die, thereby advancing geographically tailored public health insights.

Methods

Tract-level mortgage lending and demographic data were integrated with narrative cause-of-death records from the Cook County Medical Examiner. A GIS-informed machine learning framework was used to classify neighborhood context and link it to mortality patterns in Cook County, Illinois.

Using tract geometries from the American Community Survey (ACS) 2020 and integrating:

2018 racial/ethnic composition
Median household income
Shifts in mortgage lending by borrower race/ethnicity and income from HMDA
CDC Social Vulnerability Index

A standardized feature set capturing racial/ethnic composition, economic resources, and structural change in lending markets was constructed. Unsupervised machine learning methods with gap statistics were used to derive five interpretable neighborhood clusters.

Text descriptions from cause-of-death fields were cleaned, lemmatized, and harmonized into medical concepts, then summarized by cluster using frequency and tf–idf contrasts.

Key Findings

The five neighborhood clusters showed distinct mortality profiles that mirrored structural conditions:

Disinvested Black neighborhoods: Marked by violence and street-level drug toxicity
Affluent White enclaves: Reflected prescription/synthetic drug deaths and self-harm
Latino working-class areas: Exhibited mixed patterns of chronic disease, injury, and substance involvement
White-Latino transition neighborhoods: Dominated by chronic illness and degenerative diseases
Lower-income Black neighborhoods: Burdened by chronic illness, degenerative diseases, infectious causes, and substance-related mortality

Predictive Modeling Results

Penalized logistic regression models (glmnet) were used to predict cluster membership from:

Text alone
Covariates (income, rent, poverty, age, gender)
Both combined

Key Results:

Text features exhibited high sensitivity across clusters
Socioeconomic covariates improved specificity, particularly in affluent White enclaves
Combined models (text + covariates) achieved the highest overall accuracy
Performance varied by cluster: affluent White enclaves were well distinguished by socioeconomic covariates, while text features more easily identified disinvested Black and lower-income Black neighborhoods

Research Questions

RQ1: Can neighborhood structural typologies derived from mortgage lending and demographic patterns be meaningfully linked to cause-of-death narratives?

RQ2: Does narrative text from Medical Examiner death records improve prediction of neighborhood typology?

Conclusions

These results highlight how integrating structured tract-level data with unstructured death narratives uncovers not only place-based disparities in mortality risk, but also systematic differences in how death is recorded across structurally distinct neighborhoods.

Five neighborhood clusters were identified using k-means clustering on tract-level indicators. Penalized logistic regression models demonstrated that:

Combined text and covariate models achieved the highest accuracy across most neighborhood types
Narrative language varied systematically by neighborhood type, reflecting place-based inequality in death documentation

Topics: Mortality patterns, neighborhood effects, health inequities, machine learning, natural language processing, GIS, medical examiner data, Cook County

Living Here Predicts How You’ll Die: Linking Death Narratives to Neighborhood Economic Disinvestment with Unsupervised Machine Learning

Wed, 19 Nov 2025 12:00:00 +0000

Development and Spatial Validation of a Random Forest Prediction Model for Firearm-Related Injury Risk in Chicago Census Tracts

Wed, 01 Oct 2025 00:00:00 +0000

Presentation at Women in Statistics and Data Science Conference

Our presentation, Modeling Firearm Injury Risk Using Spatial Random Forests, was delivered by Dr. Gia Barboza-Salerno at the Women in Statistics and Data Science on November 12, 2025 in Cincinnati, OH.

Abstract

This research develops and validates a spatial random forest prediction model for firearm-related injury risk at the census tract level in Chicago. By incorporating spatial autocorrelation and neighborhood-level predictors, the model achieves high predictive accuracy in identifying areas at elevated risk for gun violence.

Methodology

Spatial Random Forest Approach

The spatial random forest model extends traditional random forest methods by:

Capturing Non-Linear Relationships: Modeling complex interactions between neighborhood characteristics and shooting risk
Handling Spatial Autocorrelation: Incorporating spatial structure through lag variables
Variable Importance: Identifying which neighborhood features are most predictive
Robust Predictions: Providing out-of-sample risk estimates

Data Sources

Shooting incident data from Chicago
Census tract-level socioeconomic indicators
Built environment features
Spatial lag variables to capture spillover effects

Model Validation

Cross-validation for predictive performance assessment
Spatial autocorrelation diagnostics
Comparison with traditional regression approaches

Key Findings

The predicted shooting risk map (shown above) demonstrates:

High Spatial Variation: Risk scores range from 0 to over 80 predicted incidents
Concentrated Risk: Certain neighborhoods show elevated firearm violence risk
Model Accuracy: Spatial random forest successfully captures fine-grained risk patterns
Spatial Clustering: Violence risk exhibits significant spatial autocorrelation

Applications

This spatial prediction model has important applications for:

Resource Allocation: Targeting violence prevention resources to high-risk areas
Prevention Planning: Understanding neighborhood-level drivers of gun violence
Policy Evaluation: Assessing how neighborhood changes affect violence risk
Early Warning: Detecting emerging hotspots before violence escalates

Implications

The research demonstrates that spatial random forest models can:

Improve prediction of firearm violence risk beyond traditional methods
Identify actionable neighborhood-level intervention targets
Account for complex spatial dependencies in violence patterns
Support evidence-based violence prevention strategies

Conference: Women in Statistics and Data Science Conference
Location: Cincinnati, Ohio
Date: October 1-3, 2025
Authors: Gia Barboza-Salerno, Hexin Yang, Taylor Harrington

Development and Spatial Validation of a Random Forest Prediction Model for Firearm-Related Injury Risk in Chicago Census Tracts

Wed, 01 Jan 2025 00:00:00 +0000

Plain-Language Summary

This study describes the development and validation of a machine learning model (random forest) to predict the risk of firearm-related injuries across Chicago neighborhoods. By combining multi-source data, including local crime reports, socioeconomic vulnerabilities, and environmental exposures, the model maps out areas with the highest predicted risk.

The model’s predictions were validated against actual administrative records to verify its spatial accuracy. This predictive modeling framework shows how machine learning can help public health officials and community organizations anticipate high-risk spatial hot spots and deploy community-based interventions more effectively.

Structured Topic Modeling of Child Fatal Injury Narratives Using Machine Learning with County-Level Demographic Indicators to Strengthen Prevention Strategies

Wed, 01 Jan 2025 00:00:00 +0000

Plain-Language Summary

This study applies machine learning and natural language processing to analyze child fatality reports in Pennsylvania from 2016 to 2023. Unstructured text narratives from child death investigations were combined with census data to identify recurring themes and environmental factors surrounding child neglect and traumatic injuries.

Using structured topic modeling, the researchers identified 11 distinct circumstances of death, including substance misuse, sleep-related deaths, firearms, and supervisory neglect. The analysis showed that counties with higher poverty levels and larger non-White populations experienced a higher prevalence of firearm-related and severe traumatic injuries. The findings demonstrate how combining clinical narratives with machine learning can strengthen fatality surveillance and help construct more effective, targeted prevention programs.

A Multilevel Analysis of Individual and Community Factors Associated With Case Dispositions Following Child Maltreatment Investigations

Tue, 18 Jun 2024 00:00:00 +0000

Plain-Language Summary

This study investigates how individual case characteristics and county-level social vulnerabilities influence the decisions made by child protective services (CPS) investigators when recommending services for families. Analyzing over 12,000 cases across New Mexico, the study looks at factors such as child age, abuse type, prior history, and county-level deprivation (e.g., housing and transportation deficits).

The results show that county-level characteristics significantly influence whether a family receives support services. Even when families present with similar risk levels, those living in counties with higher structural vulnerabilities are subject to different service disposition outcomes. The findings highlight the need for standardized decision-making tools to ensure equitable treatment and service distribution across different geographic regions.