If all a company is doing is collecting data and then performing basic
data access and reporting on that data (activities like standard reporting,
queries and data drill-downs at the bottom of the pyramid in Figure I), then
the questions that can be answered will be very basic. Users will be able to
tell what happened, as well as where, when, and how often, but that’s about it.
In order to create leading indicators that can answer more strategic business questions like "Why is this happening?" and "What if these trends continue?" we must move up the analytics pyramid by employing advanced analytics techniques such as statistical analysis and forecasting and extrapolation. It is only then that we can create the forward-looking leading indicators that our leaders outside of safety really want, and are starting to expect.
Davenport suggests that to answer the penultimate business question of "What will happen next?" we need to go beyond even advanced analytics and employ predictive modeling. Predictive modeling has proven to be extremely successful in returning value in other business areas like sales forecasting, customer retention, and customer upselling. And over the last few years, predictive models have been developed in safety to predict future workplace injuries so that they can be prevented.
A research team at Carnegie Mellon University (CMU) in Pittsburgh, Penn. – part of the same group that helped IBM build the Deep Blue and Watson supercomputers – used four years of real-world safety data to build computer predictive models in safety. These models were tested at overall accuracy of rates of 80 to 97 percent. With an r-squared correlation measure of .75, the research team was able to explain 75 percent of the variation in a company’s injury rate based on their safety inspection and observation data.
Talk about a leading indicator! This model, which is now in production and being used by nearly 100 companies, uses a company’s safety inspection and observation data from the last three months to predict the number and location of safety incidents over the next 30 days.
A Real-World Safety Prediction Story
On Feb. 27, 2013, a worker at a location in Hazard, Ken. lacerated his arm with a box-cutter and had to go to the hospital for stitches.
The employee was not significantly hurt, but nonetheless the company recorded
its first lost-time incident at that location in nine years. Because this
location was the safest within this business unit, there were no obvious
indicators to suggest heightened risk of injury at this plant.
However, on Feb. 1, the CMU-developed "Red Flag Prediction Model" identified this location as being at high risk of having increased safety incidents. The machine-learning computer model used the last three months of safety inspection and observation data from the Hazard location to predict that it was going to have an incident – when no other indicators were suggesting anything of the sort. The computer model saw trends and nuances in the safety inspection and observation data that allowed it to derive conclusions that were beyond the reach of traditional safety measures.
Now, keep in mind, the worker in this prediction success story still got hurt. While predicting injuries is getting easier, preventing them is as hard as it’s ever been. The Red Flag Prediction Model is still quite general in that it cannot inform a company of exactly who, when, or how someone will get hurt. But over time, as more granular data gets collected and analyzed, it is believed that these prediction models will only get more specific and accurate.
Regardless, this same company has had six locations "red flagged" by the prediction model over the last 12 months. Four of these six sites had incidents. The company is unsure if the two that did NOT have incidents were due to their active intervention to prevent injuries, or because the model simply made an incorrect prediction. Regardless, with four out of six site predictions proving accurate, this company now has a high confidence level in the predictive model and has made it a key part of their safety program. When a location is red flagged, they have a safety stand-down at that location and review their highest areas of risk. They use the model like the "check-engine" indicator in a car. It can’t tell them exactly what is going on, but it can point them in the direction of their highest risk areas.
The Top of the Pyramid
As Davenport suggests, once a company can predict what will happen in
the future, they can optimize their response to this prediction and achieve the
best outcome (the pinnacle of the analytics pyramid in Figure I). In safety,
that means predicting and then preventing workplace injuries.
The company who had the box-cutting incident already had a well-below average incident rate for their industry. They had achieved this by employing many of the traditional safety strategies discussed earlier in this article. However, to get to the "last mile," or zero incidents, they have turned to new strategies, including advanced and predictive analytics on their safety big data set. Employing technology, they have increased their safety inspection data set by 700 percent which has fueled the advanced and predictive analytics resulting in reductions across key loss metrics. They reduced their recordable incident rate by 76 percent, their lost-time incident rate by 88 percent, and their lost-work day rate by 97 percent. By employing data analytics strategies up and down the pyramid in Figure I, they estimate their return on investment from their analytics efforts at around 20,213 percent from loss category reductions alone. No wonder safety data analytics is such a hot topic!
Griffin Schultz is the general manager of Predictive Solutions Corporation, whose vision is to end death on the job, in this century. Schultz has an MBA from the Wharton School and has deep experience using technologies to solve difficult business challenges. He can be reached at gschultz@predictivesolutions.com.