NEISS-AI · open injury tracker
National estimates of injuries treated in hospital emergency departments, from the US Consumer Product Safety Commission's NEISS data. An AI text classifier reads the injury stories, so e-bike counts stay consistent across CPSC's 2024 product-code change.
E-bike emergency visits, 2025
Children (0–17) hurt on e-bikes, 2025
E-scooter emergency visits, 2025
Estimated US visits each year with 95% confidence bands. Regular bicycles (about 400,000 a year, roughly flat) are left out so the newer devices can be seen; add them in Explore the data.
Hollow points are estimates CPSC would call potentially unstable (fewer than 20 records, under 1,200, or coefficient of variation above 33%).
Is the rise just a change in coding? No. CPSC gave e-bikes their own product code in 2024, and both 2024 and 2025 use that separate code, so the 2024–2025 rise is not a coding effect. For 2019–2023, when e-bikes shared a code with mopeds, NEISS-AI uses its classifier so every year is counted the same way.
Estimated US emergency department visits in each month of the chosen year, with 95% confidence bands. Switch to "Share of the year" to compare seasons across devices, including regular bicycles.
Monthly estimates rest on fewer records than yearly ones, so their ranges are wider. In a year when e-bike use grows quickly, later months also tend to be higher for that reason alone.
Some "e-bikes" in hospital records are really electric motorcycles (e-motos, such as Sur-Ron or Talaria) or e-bikes ridden well above legal e-bike speeds. NEISS-AI reads each e-bike story and sorts it into three groups. Likely e-motos are about 2% of e-bike visits every year, and that share has not grown, so the e-motos named in hospital stories do not explain the rise in e-bike injuries.
How the groups are made. Likely e-moto: the story uses e-moto wording (electric motorcycle or dirt bike, gas- or engine-powered, an e-moto brand) or gives the rider's own speed above 28 mph, the top assisted speed of a legal Class 3 e-bike. Unclear: wording that could mean either ("motorized bike", "motor bike" with no high speed, "dirt bike or e-bike"), a speed range crossing 28 mph, a downhill speed above 28 mph, or "fast" with no speed given. Diagnosis coding phrases are ignored. Regular e-bike: everything else. Speeds of other vehicles ("struck by a car going 30 mph") are ignored. Accuracy. The rule was built on 320 stories, then locked and tested once on 240 new stories from 2019–2025 that it had never seen. The test stories were labelled blind against written definitions, without seeing the rule. Across all e-bike stories the rule puts 98% in the right group (95% range 97–99%), and it estimates the share of likely e-motos almost exactly (2.0% by the rule vs 2.0% by the blind labels). Single "likely e-moto" calls are right about 3 times in 4: most misses are "motorized bicycles" ridden over 28 mph, which the definitions count as unclear. Full details: methods note.
Read with care. A story names an e-moto only when the writer says so, so "likely e-moto" is a lower bound. CPSC blacks out brand names in the stories ("***"), so most likely e-motos are found from the rider's stated speed; an e-moto described only as an "e-bike" cannot be found. The e-bike class (1, 2 or 3) is almost never written in the story, so classes cannot be counted. Earlier years have more unclear cases because stories then often said "motorized bicycle" rather than "e-bike".
E-bike visits by age group. Children's visits rose sevenfold from 2022 to 2025.
Share of injured riders admitted, transferred or held, 2019–2025. For children, e-bike injuries are more serious than bicycle injuries.
Ask about devices, ages, injuries, outcomes and years. The answer is computed from the NEISS records in your browser, with a 95% confidence interval.
Pick devices and filters. Every number is a national estimate with a 95% confidence interval.
Every NEISS record has a short story written by hospital staff. Search them (for example battery, hit by car, helmet, delivery) to see national estimates and real examples. All words must appear.
Three checks, all repeatable with the open code.
1. The classifier tells e-bikes and mopeds apart almost perfectly. It was tested on 2024–2025 injury stories it had not seen during training (five-fold cross-validation). Its AUC was 0.995: shown one real e-bike story and one real moped story, it ranks the e-bike story higher 99.5 times out of 100. (0.5 would be a coin flip; 1.0 is perfect.)
2. The national totals match CPSC's own published figures. The pre-2024 stories have no answer key, so the test is the total: NEISS-AI's national e-bike estimates for 2022–2024 are within 5% of the figures CPSC published in its 2025 micromobility report.
| Year | NEISS-AI | CPSC published | Difference |
|---|---|---|---|
| 2022 | 28,000 | 26,800 | +4.4% |
| 2023 | 37,900 | 37,600 | +0.7% |
| 2024 | 59,300 | 59,200 | +0.3% |
E-bike emergency department visits, national estimates, rounded to the nearest 100; differences use unrounded NEISS-AI estimates. CPSC source: Micromobility Products-Related Deaths, Injuries, and Hazard Patterns: 2017–2024 (CPSC, 2025).
3. Car involvement and helmet use are checked on stories the rules never saw. NEISS does not code whether a motor vehicle was involved or whether the rider wore a helmet, but the injury story often says so. Since version 1.4, NEISS-AI reads both from the story. On a test set of 300 random stories (2019–2025) that were not used to build the rules, it was right for motor vehicle involvement in 98% of stories (it found 44 of 47 real cases, and 44 of its 46 flags were correct) and for helmet status in 300 of 300. Helmet status has three answers: worn, not worn, or not mentioned; "not mentioned" (about 4 in 5 stories) is never counted as "no helmet". The reference answers came from an independent AI reviewer that never saw the rules, with disagreements settled against the same written definitions; they were not made by a human reviewer.
Limits: stories are short, so a few are uncertain; NEISS covers emergency department visits only; counts are not risk rates, because the number of riders is not known.
Free and open source (MIT licence). Download the NEISS annual files from CPSC into a folder called data, then:
pip install neissai
neissai micromobility --years 2019-2025 --data-dir data --out-dir results
neissai classify "14YOM FELL OFF HIS ELECTRIC BIKE GOING 20MPH"One-page summaries with citations, ready to print or attach to a report.
If NEISS-AI helps your research, teaching or reporting, please cite the software and the method paper.
Kumar S. NEISS-AI: AI-assisted national injury estimates from NEISS emergency department narratives. Version 1.4.0 [software]. Zenodo; 2026. https://doi.org/10.5281/zenodo.23550272
Kumar S. E-bike injuries among US children surged in 2025: national estimates from emergency department reports, 2019–2025, using machine learning to read injury narratives [preprint]. Zenodo; 2026. https://doi.org/10.5281/zenodo.23354066
Kumar S. NEISS-AI device labels for US micromobility injuries, 2019–2025 [dataset]. Version 1.1. Zenodo; 2026. https://doi.org/10.5281/zenodo.23550382
For reference managers: Download BibTeX Download RIS (EndNote, Zotero, Mendeley)
I would love to hear how you use this tool, for research, a class, a news story, a safety program or policy work. A two-line note helps me improve it and keep it free.
Limits: NEISS counts emergency visits, not riders or miles, so these are counts, not risk rates. NEISS has no state or city information. Helmet use and motor vehicle involvement are only what the story mentions; about 4 in 5 stories do not mention a helmet. Early e-bike estimates (2019–2021) rest on small samples. The question box understands common phrases, not every possible question; check the "Understood as" line under each answer.