Safety × Research note
Where serious construction injuries come from, and what incident reports can tell you
Falls get most of the attention in safety plans, and rightly so. But the hazards that put construction workers in hospital are not always the ones plans are written around, and the detail that explains them sits unread in incident reports. What the evidence shows, with a ten-year case study of 19,021 severe-injury reports.
Key takeaways
- Falls are the most common severe injury in construction, and in the case study one in four falls with a recorded height was from under six feet. Ladders and low platforms matter, not only roofs.
- Caught-in and crushing accidents are rarer but cause most amputations: 63% of them involve one.
- The hazard mix changes sharply by trade and by season. 80% of heat-stress injuries happen from June to August.
- Incident narratives can be coded consistently by a simple, explainable model: 84% agreement with OSHA’s coders against 55% for a keyword list.
- Coding rules change. OSHA’s January 2024 manual change moved the caught-in/struck-by line, so any model needs monitoring by cause.
01What most safety data misses
Most contractors track incident counts, lost-time injury rates and near misses. Those numbers are useful, but they are dominated by frequent, less severe events, and they say little about why people are hurt. The detail that explains an injury (the task, the equipment, the height, the sequence of events) sits in the free-text narrative of the incident report.
That text is rarely analysed. Coding each report consistently by cause takes trained people and time, so in most firms the narratives are read once, filed and never compared. The result is a safety plan built on the hazards everyone expects, rather than on the pattern the firm’s own reports would show.
The detail that explains an injury sits in the narrative, and the narrative is rarely read twice.
02The most serious hazards are well known
Fatality statistics tell a consistent story. In detailed national fatality data such as the Bureau of Labor Statistics’ census, construction and extraction workers account for about one in five of all workplace deaths, and falls, slips and trips are their leading cause of death[5]. Roofing work alone accounts for about a quarter of construction’s fatal falls[6].
Enforcement records point the same way. Fall protection is routinely the most frequently cited construction safety standard; OSHA reported it at the top of its list for the fourteenth year running in 2024[4]. The main hazards are not a secret. What varies between firms is whether their own data shows where those hazards are hurting their people.
03Case study: Construction Safety Intelligence
From my portfolio · project 01 of 15
Real public data (OSHA)- Problem
- Contractors’ injury reports stay as unread text because most firms have no trained coders.
- Decision supported
- Where to focus safety effort, by hazard and by trade.
- Data
- 19,021 OSHA severe-injury reports from construction, Jan 2015 – Nov 2025.
- Method
- Keyword baseline, linear text models, CNN, BiLSTM and a blind LLM; trained on earlier years, tested on 2024–2025.
- Result
- 84% agreement with OSHA’s coders (keywords: 55%); 89% where model and LLM agree.
- Limits
- Severe injuries only, from one regulator’s jurisdiction; a 2024 coding-manual change needs monitoring.
04The data: ten years of severe-injury reports
Since 1 January 2015, employers under OSHA’s federal jurisdiction must report any work-related in-patient hospitalisation, amputation or loss of an eye within 24 hours[2]. Each report includes a short free-text narrative of what happened, and OSHA’s staff code it by event, source and body part. OSHA publishes the full file on its Severe Injury Reports page[3].
I downloaded the January 2015 to November 2025 file (105,996 reports across all industries) and kept the 19,021 reports from construction (NAICS sector 23). The cleaning mattered more than the modelling. The file holds 348 event codes stored at two, three or four digits, from more than one version of OSHA’s coding manual, where the same prefix can mean different things. I grouped them into nine cause groups from their titles, removed employer names, addresses and coordinates, and set aside 381 reports with no usable cause[1].
05Where serious injuries come from
Of the 18,640 reports with a known cause, falls to a lower level are the largest group by a distance. But the share of each cause that ends in an amputation tells a different story about severity[1].
Figure
Severe construction injuries by cause, Jan 2015 – Nov 2025 (case study)
| Cause | Reports | With an amputation |
|---|---|---|
| Fall to lower level | 6,806 | 0.4% |
| Struck by or against object | 4,706 | 24.8% |
| Caught in or crushed | 2,285 | 62.8% |
| Vehicle or mobile equipment | 1,429 | 6.7% |
| Electrical | 959 | 2.3% |
| Slip, trip or same-level fall | 881 | 2.7% |
| Heat stress | 605 | 0% |
| Fire, explosion or burn | 497 | 1.6% |
Falls put the most people in hospital; caught-in accidents take the most fingers and hands. A safety plan that ranks hazards only by frequency will under-weight the machinery, pinch points and rotating parts behind the second group.
06Three findings a safety manager can use
1. Low falls are a large share of serious falls
One in four falls with a recorded height band was from under six feet[1]. Fall-protection effort naturally goes to roofs and leading edges, but ladders, trestles and low platforms produce a steady stream of hospitalisations. When I asked a language model to extract the equipment from 2024–2025 narratives, ladders were involved in 407 falls (median fall 8.5 ft), roofs in 193 (16 ft) and scaffolds in 111 (10.5 ft)[1].
2. The hazard mix depends on the trade
Over half of injuries in foundation, structure and roofing contractors, and in finishing trades, are falls. Utility contractors have the most struck-by and caught-in injuries, and electrical, plumbing and HVAC contractors account for 57% of electrical injuries[1]. A company-wide top-ten hazard list hides this; a per-trade view does not.
3. Heat is seasonal and regional
80% of heat-stress injuries in the file happened from June to August, mostly in Texas and Florida[1]. That makes heat one of the few hazards you can plan for by calendar: water, shade, rest cycles and acclimatisation for new starters, scheduled before the season rather than after the first incident.
Falls put the most people in hospital. Caught-in accidents take the most fingers and hands. A plan ranked only by frequency misses the second.
07Reading incident text at scale
Most contractors don’t employ trained coders, so their own incident reports stay as unread text. I tested whether a model could code the cause of an injury from the narrative as consistently as OSHA’s coders, training on 2015–2022, tuning on 2023 and testing on 2024–2025 so the model is always judged on later years than it learned from[1].
Figure
Agreement with OSHA’s coders on 3,204 test reports (2024 – Nov 2025)
Three lessons carry over to any contractor:
- Simple models are enough and explainable. Narratives are short (median 190 characters), and a linear model with one weight per word beat neural networks. The strongest words read like a safety manager’s vocabulary: “pinched”, “between”, “amputation” for caught-in; “backed”, “forklift” for vehicles.
- Two readers make a review queue. Where the trained model and a language model agree (85% of reports), they match OSHA 89% of the time; only the 15% where they disagree need a person.
- Labels drift. OSHA changed its coding manual in January 2024 and redrew the caught-in/struck-by line. The narratives didn’t change but the codes did, and caught-in F1 fell from 0.78 to 0.59. Any production model needs monthly monitoring by cause.
08How strong is the evidence?
Not every finding in this note rests on the same kind of evidence. This is how I would weigh each one before acting on it.
| Finding | Evidence | Strength | Main caveat |
|---|---|---|---|
| Falls are the largest cause of severe injuries | 19,021 official OSHA reports over ten years | Strong | One regulator’s jurisdiction; severe injuries only |
| Caught-in accidents drive most amputations | Same data, amputation recorded per report | Strong | Coding line with struck-by moved in 2024 |
| Falls, slips and trips lead construction fatalities | BLS Census of Fatal Occupational Injuries | Strong | Counts, not rates per hour worked |
| A simple model codes causes about as well as coders agree | Time-split test on 3,204 later reports | Moderate | One country’s reports; re-test on your own |
09What to do with this on your projects
- Rank hazards by severity as well as frequency. Track amputations and hospitalisations separately; caught-in risks rise up the list.
- Treat ladders and low work as fall hazards. Include step ladders, trestles and platforms under six feet in fall-protection planning and inspections.
- Split the dashboard by trade. A roofing crew and a utility crew need different top-three hazards.
- Put heat on the calendar. Plan water, shade and acclimatisation before June, not after the first incident.
- Code your own incident text consistently. A small, explainable model plus a person reviewing disagreements turns narratives into trend data.
- Monitor the codes, not just the model. When definitions change, re-measure accuracy by cause before trusting the trend line.
NotesSources
- Iwale, A. (2026). Construction Safety Intelligence: analysis of OSHA Severe Injury Reports, January 2015 – November 2025. GitHub.
- OSHA. 29 CFR 1904.39 — Reporting fatalities, hospitalizations, amputations, and losses of an eye.
- OSHA. Severe Injury Reports (data set and dashboard).
- OSHA. Top 10 Most Frequently Cited Standards, fiscal year 2024.
- U.S. Bureau of Labor Statistics (2025). Fatal work injuries fell in 2023. The Economics Daily.
- U.S. Bureau of Labor Statistics (2025). Fatal falls in the construction industry in 2023. The Economics Daily.
Figures are quoted from the sources above as published; where a source reports a range or a survey estimate, it is described that way. Results from my own projects say whether they use real public data or synthetic data.