Predictive Safety Analytics: A Practical Guide

Most safety programs look backwards. They track injuries, count lost-time incidents, and review what went wrong after someone got hurt. That's useful. But it means you're always reacting.

Predictive safety analytics flips that model. Instead of asking "what happened?", it asks "what's most likely to happen next — and where?"

This guide explains how it works, what data it needs, and how to build a scoring system you can actually use.

The honest limitation you should know upfront

Before getting into formulas and frameworks, here's something the software vendors won't tell you: there is no scientifically validated formula that predicts a specific incident on a specific date.

A 2026 review published in OHS Canada put it plainly — what we call leading indicators are measures of activity and control effectiveness, not forecasting tools. The lesson from the research is that you can score relative risk across your sites and work areas, and rank what needs attention most. That's not a small thing. But it's different from prediction in the strict sense.

So the right way to think about this: predictive analytics gives you a ranked list of where risk is highest right now. That tells you where to direct inspections, training, and interventions. Done consistently, that reduces incidents. The Campbell Institute proved this with a dataset of over 112 million safety observations across 15,000 worksites — inspection and observation data strongly predicted future incidents, and organisations that acted on it had fewer of them.

Leading vs lagging: what actually matters

Most safety teams collect lagging indicators by default. Injury rates, LTIFR, TRIFR, near miss counts after the fact. These are worth tracking, but they have a ceiling problem: as your workplace gets safer, you have fewer incidents to learn from. One organisation studied by Predictive Solutions saw its incident rate drop 95% over a year, leaving just 20 lagging data points to analyse — while they had over 8,000 inspection data points sitting unused.

Leading indicators are the forward-facing signals. They include:

• Near miss frequency (rate per 100 workers per month)
• Inspection completion rate
• Corrective action close-out rate
• Safety observation volume
• Training compliance rate
• Overdue hazard register items

The key insight is that these signals degrade before incidents occur. Inspection completion drops. Near misses stop being reported (which usually means the culture has shifted, not that conditions improved). Corrective actions pile up. If you're watching these numbers, you'll see the deterioration before someone gets hurt.

A practical risk scoring formula

The most useful framework to come out of recent research is the Weighted Composite Score (WCS), introduced in a 2025 study published in MDPI's occupational health journal. It improves on the classic Severity × Probability risk matrix by adding variables that the standard model misses.

Here's a practical version adapted for WHS use:

Incident Risk Score (IRS)

IRS = (w1 × S) + (w2 × P) + (w3 × F) + (w4 × L) + (w5 × C) + (w6 × H)

IRS range: 1–5. Above 3.5 = high risk. Above 4.2 = critical.

A few notes on using this:

The weights above are a reasonable starting point, derived from the MDPI research and the FMEA+ method (a peer-reviewed framework validated in the steel industry, where occurrence, severity, and detectability were weighted at 0.337, 0.348, and 0.315 respectively). But they aren't universal. The right weights for your business depend on your industry, your workforce, and your incident history. The formula's real value is in applying it consistently and calibrating it over time against your own data.

The L variable (leading indicators) deserves extra attention because it's the most actionable. You can't change a worker's prior incident history. You can change whether inspections are being completed and whether near misses are getting reported and closed out.

What the context variable captures

The C variable is often what tips a borderline situation into an actual incident. It covers:

• Fatigue and overtime — hours above baseline for key workers
• Crew change rate — new or unfamiliar workers on a task
• Production pressure — deadline-driven workload spikes
• Environmental conditions — heat, weather, poor visibility
• Organisational change — restructures, new procedures, system changes

These factors don't create hazards on their own, but they amplify existing ones. A hazard with a baseline IRS of 3.2 might score 4.0 when a crew changeover coincides with a deadline push and two weeks of heat.

How to apply this in practice

Run the IRS calculation per hazard or per work area, on a rolling 30-day window. The output isn't a prediction — it's a ranked list. Your top five scores tell you where to focus your next round of inspections and conversations.

The process looks like this:

1. Score each active hazard across the six variables
2. Rank by IRS
3. Review anything above 3.5
4. Act on the leading indicator inputs — not just the score
5. Recalculate after interventions to see if the score moves

The last step is important. If you run inspections, close out corrective actions, and retrain workers but the score doesn't drop, something is being missed.

What data you actually need to get started

You don't need sophisticated AI to start scoring risk this way. You need clean, consistent data. Specifically:

• A hazard register with severity and exposure frequency recorded
• Inspection records with completion rates tracked over time
• Near miss reports with close-out dates
• Training records by worker and task
• Incident history by site and work type

Most WHS software already collects this. The gap for most organisations isn't data collection — it's that the data sits in separate places and no one is calculating a combined score from it.

The Bird's Pyramid connection

Frank Bird's safety pyramid (from his 1966 study of 1.7 million accidents across 297 companies) established that for every serious injury, there are roughly 10 minor injuries, 30 property damage incidents, and 600 near misses. The pyramid isn't a precise formula — the ratios vary by industry — but the principle holds: near misses are the most abundant signal you have, and they precede serious incidents in observable patterns.

This is why the L variable in the IRS gives near miss frequency meaningful weight. A site that reports zero near misses isn't necessarily safer than one reporting ten a month. It may just have a weaker reporting culture — which is itself a risk signal.

How Safety Space supports this approach

Safety Space collects the data that feeds this model — hazard registers, inspection records, near miss reports, corrective actions, and training compliance — and surfaces it in a format that lets you identify where risk is concentrating before an incident occurs.

If you want to see how this works against your own site data, a 10-minute demo is the fastest way to assess it.

Further reading

• Predictive Analytics in Workplace Safety: Four Safety Truths — Campbell Institute / Predictive Solutions & Carnegie Mellon University (112 million observations)
• A Risk Management Approach in OHS Based on the Integration of a Weighted Composite Score — MDPI Occupational Health, October 2025
• The Myth of Predictive Safety Indicators — OHS Canada, April 2026
• Introducing FMEA+ Method for Comprehensive Safety Risk Assessment — PubMed Central
• Advancing Safety Analytics: A Diagnostic Framework for Assessing System Readiness — PubMed Central
• Accident Triangle (Bird's Pyramid) — Wikipedia overview with source history