The Binder Couldn't See It

Every safety program has a cabinet. Inspection sheets, near-miss cards, hazard assessments, incident reports, years of them. Somewhere in those pages is the pattern that would have told you which site, which task and which month was heading for trouble. Nobody found it, because nobody had time to read four thousand pages and keep a tally.

The problem was never a shortage of data. It was the cost of reading it. Researchers at the University of Colorado put it plainly when they set out to mine construction injury reports: because there was no clear method and manual content analysis at scale was so expensive, "these valuable data have yet to be extracted and leveraged."¹ That sentence describes most safety departments in the paper era. The information was collected faithfully, filed, and left there.

What a computer can read in an afternoon

The same team built a natural language processing tool that scans free-text injury reports for 101 attributes and outcomes, the kind of detail a safety professional would pull out by hand, and reported over 95% accuracy.¹ They then trained machine learning models on those attributes and could predict the injury type, the energy involved and the body part hurt with real skill. Their conclusion is the one every field supervisor suspects: the results suggest "injuries do not occur at random."²

A follow-up study grew the dataset to more than 90,000 reports and, unlike the original work, predicted injury severity well.³ A 2022 systematic review found the same techniques being used across industries to classify accident types, identify causal factors and predict injuries, and noted that the deep learning side of the field is still young.⁴ The direction is clear even if the tools are still maturing.

“The information was collected faithfully, filed, and left there.”

Leading, lagging, and the gap between them

OSHA describes leading indicators as proactive measures that can "reveal potential problems in a safety and health program," and lagging indicators as measures of "events that occurred in the past," such as injury rates.⁵ Canada's CCOHS draws the same line: lagging indicators are "a record of what has happened."⁶ Interestingly, CCOHS files near misses under lagging, because a near miss has already happened. What leads is everything you do before it: the inspections, the observations, the hazard assessments, and whether the fixes held.

Lagging numbers are easy to count and arrive too late. Leading numbers arrive in time but used to be impossible to tally across dozens of crews and hundreds of forms each week. That is the gap software closes. When a hazard assessment is filed on a phone instead of a clipboard, it is counted the moment it is filed, by site, by hazard and by crew, without anyone doing the arithmetic.

From a binder to an early warning

Counting is only half of it. The other half is knowing when a number has actually changed, rather than wobbling the way numbers do. Quality engineers solved this a century ago with the control chart: plot the measure every week, mark its usual level, and draw an upper limit about three standard deviations above. A point above the limit is a real signal, not noise. Public health researchers applied exactly this to injuries at 49 US Army installations and found four running above their upper limit, three below, and the rest moving around their usual level.⁷ Four places to look out of forty-nine is a manageable Monday.

This is the work FieldTec was built for. Crews file their hazard assessments, observations and inspections from the field in a few taps, and the people running the business see what is rising, and where, while there is still time to do something about it.

Inspection data turns out to be just as telling. Research by Predictive Solutions and Carnegie Mellon University on large volumes of inspection records reported predictive models with accuracy in the range of 80 to 97 percent, and summarized four findings that have stuck with the industry ever since.⁸ It is vendor research rather than a peer-reviewed study, so treat the numbers with care. The findings themselves match what experienced safety people see:

Where the pattern lives

The most useful view for a monthly review is often the simplest: sites down one side, hazard types across the top, and the count in each square. The hot squares are where to spend next month's attention. On paper, building this meant a clerk, a spreadsheet and a week. Filed digitally, it is there every morning.

In FieldTec that view is simply there each week, by site and by hazard, for the person who has to decide where the next inspection goes.

Two cautions before you trust the chart

First, more of the small stuff does not mean more of the serious stuff. Fred Manuele's review of Heinrich's famous accident ratios concluded that the data behind them is "unsupportable," and that reducing how often incidents happen "does not necessarily achieve an equivalent reduction in injury severity."⁹ Trend the events with serious potential on their own. A falling count of paper cuts says nothing about the next fall from height.

Second, a trend is only as good as the reporting under it. A study of 38 construction contractors found that where the safety climate was poor, 81% of eligible injuries went unreported, compared with 47% where the climate was positive.¹⁰ A quiet chart can mean a safe site, or a crew that stopped telling you. That is why participation, the share of workers who report anything at all, belongs beside every trend.

In 2023, 1,057 workplace fatalities were recorded in Canada, along with 274,022 accepted lost-time claims.¹¹ Behind most of those numbers is a trail of earlier signals: an observation, a finding, a hazard ticked on a form. For most of the history of this profession, those signals were written down and never read together. They can be now. In Part 2 we look at the catch: a trend is only as honest as the forms underneath it.