Today at ENR’s FutureTech conference in San Francisco, Jit Kee Chin, EVP and Chief Data Officer at Suffolk, announced the results of 6 months of collaboration between our companies. We think it’s groundbreaking and hope you will too.
The collaboration started with a simple question - can we move from automatically “observing” safety risks in project photos to predicting if they will occur at all? These “predictive models” have been used in other industries for some time. Predictive models consume a variety of data over time and learn relationships between the various inputs and what happens as a result. For example, in 2009, Google built a predictive model built from years of employee reviews, promotion, and pay histories to determine which employees were most likely to quit. Other companies followed suit to improve employee retention with the bank Credit Suisse proactively offering new roles to those employees who have been identified as likely to quit. In our case, we worked with Suffolk to see if we could build a predictive model that can tell if a safety incident will happen next week on a jobsite.
Our results? After analyzing over 10 years worth of data, including 700,000 photos reviewed by our AI engine for potential safety risks (example below) and other project related factors, we built a model that predicts when safety incidents will happen: an "incident early warning" system for construction jobsites.
Missing PPE tag “No Gloves” flagged in data from Procore; Photo courtesy of Suffolk, 2018
Example of early safety incident warning system from Smartvid.io showing estimated
risk of a safety incident from predictive analytics, FutureTech 2018 presentation.
The incident early warning system will suggest if an incident is likely to occur next week by analyzing all of the data from all the prior weeks on the project. You can choose how “sensitive” you want the warning system to be. For example, if you want certainty in the alerts, the early warning system will notify you only if it is very confident an incident will occur, being correct roughly 80% of the time and predicting 1 in 5 incidents. If want to be alerted more frequently and catch more potential incidents, you can get notified of 2 in 5 incidents with correct predictions occurring 2 out of 3 times.
Let’s make some conservative assumptions and see what the potential impact might be. Just knowing an incident may happen doesn’t mean a team can prevent it. That said, if a warning of elevated risk resulted in additional on site measures (e.g., tool box talks, visits from safety resources, additional emphasis in pre-task planning) and that resulted in only 25% of predicted incidents being avoided, this translates into 40-100 incidents avoided per year for a company with 50 ongoing projects. At a cost of roughly $36,000/incident in 2018 dollars (NIH Study, Costs of Occupational Injuries in Construction in the United States), that’s somewhere between $1.4M and $3.6M in savings per year.
It’s time these predictive models were used to do more than retain employees - they can be used to reduce risk in construction. Reach out to us at firstname.lastname@example.org to learn more.