Layered Labs

DatasetsMarch 2026

NEISS Injury Data: twenty years of US emergency injuries in one file

0
emergency department visits
0
years, 2005 to 2024
0
file instead of twenty

1Summary

NEISS Injury Data is 7,326,429 emergency department visits for injuries in the United States, from 2005 to 2024, merged from twenty yearly files into one file. It comes from the National Electronic Injury Surveillance System, run by the US Consumer Product Safety Commission, and keeps the sample weights needed for national estimates.

2Motivation

Twenty yearly files, merged into one.

Studying injuries over time with NEISS meant downloading twenty separate yearly files, each with slightly different column encodings and formats, and writing code to join them. One file with a year column removes that step: a question about twenty years of injuries becomes one query.

3Composition

The columns of each visit. The weight turns rows into national estimates.

One row per emergency department visit, with the patient's age and sex, the body part, diagnosis, outcome, the product or activity involved, a short free-text narrative, the sample weight, and the year.

ColumnWhat it holds
Treatment_DateDate of the emergency visit
Age, SexThe patient's age in years, and sex
Body_Part, DiagnosisNEISS codes for the injury
DispositionOutcome, such as treated and released, or admitted
Product_1The product or activity involved
Narrative_1A short free-text description
WeightSample weight for national estimates
yearYear of the visit, added when merging
The main columns. Codes are read with the official NEISS coding manual.

4Collection

From the Commission's yearly releases to one file.

NEISS records injuries treated in a probability sample of about 100 US hospital emergency departments, chosen so that weighted counts give national estimates. The Commission publishes the records once a year. This dataset takes every yearly file from 2005 to 2024, aligns their columns, adds a year column, and writes one Parquet file.

5Cleaning and labels

Codes are stored as numbers; the NEISS coding manual says what each means.

The records are as the Commission published them: de-identified, and not re-coded. Diagnosis, body part and product are numeric codes, which need the official NEISS coding manual to read. The narratives are free text and vary in format.

6Uses

Illustration. Each sampled visit stands for many visits nationwide; sum the weights, not the rows.

For research on injury trends, product safety, sports and recreation injuries, and emergency department use, and for policy analysis and journalism.

Each row is one sampled visit, not one injury nationwide. National estimates come from summing the Weight column, never from counting rows:

import pandas as pd
df = pd.read_parquet("hf://datasets/Layered-Labs/neiss-injury-data/neiss_all.parquet")
picks = df[(df["Product_1"] == 3235) & (df["year"] >= 2015)]  # pickleball
print(f"{picks['Weight'].sum():,.0f} estimated ED visits, 2015 to 2024")

It covers emergency visits only: injuries treated elsewhere, and deaths without an emergency visit, are not in it. Race and ethnicity coding changed over the years, so compare those fields across years with care.

7Distribution

One file, one call.

On Hugging Face as Layered-Labs/neiss-injury-data, with the code that builds it on GitHub at Layered-Labs/neiss-injury-data. The records are public and de-identified, from the US Consumer Product Safety Commission.

8Maintenance

The twenty years the file covers.

Layered Labs maintains the dataset. Questions and corrections go to its Community tab on Hugging Face or to the GitHub repository.

Living draft, updated November 2026. This page can change before formal publication.