NLP | ISSUES | The Ohio State University

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 .

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

How to Analyze Caselaw Using CourtListener's API in R

Fri, 01 Aug 2025 00:00:00 +0000

I previously showed how to pull Indian Child Welfare Act (ICWA) cases using the Caselaw Access Project (Harvard Law). That API has since changed. This post updates the workflow using the Free Law Project’s CourtListener, which now partners with Harvard to make this data freely available.

Finding the Cases

Start at the CourtListener website and search for “Indian Child Welfare Act” (or "icwa"). You’ll see a page of opinion results.

Click Select Jurisdiction → State. Clear all, then select Ohio Supreme Court, Ohio Court of Appeals, and Ohio Court of Claims. This gives you a quick check on the expected volume to validate the code’s output below.

The site reported 36 cases from Ohio. Note that the exact number may change over time as new opinions are added or metadata is corrected.

Getting the Data into R

The Search API returns clean JSON results by querying for opinions that contain the exact phrase "indian child welfare act", restricting to published opinions, ordering by relevance, and (optionally) limiting results to three Ohio courts.

library(httr2)
library(purrr)
library(dplyr)
library(tibble)
library(tidyr)
library(stringr)

SEARCH_BASE <- "https://www.courtlistener.com/api/rest/v4/search/"
UA <- "R/httr2 (your-email-address) – ICWA metadata collection"

query_params <- list(
 q = '"indian child welfare act"',
 type = "o",
 order_by = "score desc",
 stat_Published = "on",
 court = "ohio ohioctapp ohioctcl" # optional: specify court
)

A few helper functions keep the parsing tidy:

pull_chr <- function(x, ...) {
 v <- purrr::pluck(x, ..., .default = NA)
 if (is.null(v) || length(v) == 0) return(NA_character_)
 as.character(v)
}

pull_num <- function(x, ...) {
 v <- purrr::pluck(x, ..., .default = NA_real_)
 suppressWarnings(as.numeric(v))
}

get_json <- function(url) {
 resp <- request(url) |> req_user_agent(UA) |> req_url_query() |> req_perform()
 resp_check_status(resp)
 resp_body_json(resp, simplifyVector = FALSE)
}

This function handles pagination automatically, following the next cursor until all pages are exhausted:

fetch_search_pages <- function() {
 req <- request(SEARCH_BASE) |>
 req_user_agent(UA) |>
 req_url_query(!!!query_params)

 dat <- req |> req_perform() |>
 (\(r){ resp_check_status(r); resp_body_json(r, simplifyVector = FALSE) })()
 pages <- list(dat)

 while (!is.null(dat$`next`)) {
 dat <- request(dat$`next`) |> req_user_agent(UA) |> req_perform() |>
 (\(r){ resp_check_status(r); resp_body_json(r, simplifyVector = FALSE) })()
 pages <- append(pages, list(dat))
 }
 pages
}

pages <- fetch_search_pages()
results <- pages |> map("results") |> flatten()

results is a list of per-case records from the API. Each element includes the case name, filing date, court identifiers, citation strings, a public URL where the caselaw text is available, and other metadata.

For Ohio, this produced ~35 unique case names versus 36 displayed on the site. Minor discrepancies happen because search hits can include multiple opinions or sibling opinions per matter, and because the index is updated over time.

Downloading the Full Text

Each opinion has an API endpoint with plain-text and HTML fields. A small loop can request the text for each case URL, with a short pause and retry between requests, and save one .txt file per opinion:

download_opinion_text <- function(url, out_dir = "opinions") {
 dir.create(out_dir, showWarnings = FALSE)
 slug <- basename(url)
 path <- file.path(out_dir, paste0(slug, ".txt"))

 resp <- request(url) |> req_user_agent(UA) |> req_perform()
 resp_check_status(resp)
 body <- resp_body_json(resp, simplifyVector = FALSE)

 text <- body$plain_text %||% body$html_with_citations %||% ""
 writeLines(text, path)
 Sys.sleep(0.5) # be considerate of rate limits
 invisible(path)
}

walk(map_chr(results, "absolute_url"), download_opinion_text)

Here are the Ohio ICWA cases downloaded to a local folder:

Visualizing the Results

You can use ggplot2 to visualize the results. Below are two examples: filings by year and volume by court.

What You Can Do Next

After downloading all data — including caselaw text and metadata — you can:

Natural Language Processing — examine sentiment or common words across opinions
Topic Modeling — explore latent themes in ICWA cases using LDA or STM
Map filings by jurisdiction — visualize geographic patterns of ICWA litigation
Integrate with other sources — link to child welfare administrative data or Census indicators

Remember to be considerate of rate limits and to include a User Agent string with your contact information in every request.

The complete workflow is available on my GitHub.

Happy Coding!