Safety Leadership: Using artificial intelligence to identify serious injury and fatality potential

Editor’s Note: Achieving and sustaining an injury-free workplace demands strong leadership. In this monthly column, experts from global consulting firm DEKRA share their point of view on what leaders need to know to guide their organizations to safety excellence.

Until recently, the best methodology for classifying serious injury and fatality potential was a decision tree. After each incident, a series of questions is posed. If any of the questions are answered affirmatively, the incident is deemed to have SIFp.

Decision trees work, but we have found that many organizations struggle to fully implement the system. They’re also difficult to sustain. As a result, some organizations rely on one person making the determination, which creates a bottleneck.

For other organizations, the understanding of the decision tree methodology is eroded, and people are inconsistent in their designation.

And finally, some organizations abandon the decision tree approach altogether, and people are asked to simply rate the incident as having SIFp or not, making subjectivity an issue.

After witnessing the challenges organizations face with SIFp classification, we explored how artificial intelligence could be a solution instead.

Case narratives and natural language processing

Natural language processing is how computers understand both written and spoken text. Natural language processing techniques are already widely used in our everyday lives, from search algorithms, spelling and grammar checking software, to language translation.

At DEKRA, we focused on the quickly evolving field of transformer neural networks. This technology was first available in 2018 by researchers from Google (Devlin et al. 2018). We were interested in fine-tuning the neural nets on our incident database to improve the algorithm’s understanding of incident data. This approach uses incident descriptions to predict SIFp.

What we have found is that, under the right circumstances, AI can achieve an acceptable level of accuracy in classifying SIFp based on an analysis of case narratives. However, the right circumstances are rare.

Challenges to AI

One important element of a SIF prevention program is the immediate classification post incident of SIFp. This classification needs to be made for all types of cases, including first aid and near misses. To make this classification requires case narratives.

However, we found that some organizations only require reporting or investigations for medical treatment incidents. Few of the organizations require investigations for near-miss incidents. Still, with others, if the incident isn’t recordable, the only information required to be provided is the injury type.

For AI to be trained, the case narrative must contain information about the exposure that created the vulnerability and a minimum set of essential facts. For example, the narrative might say the employee fell off a ladder (the exposure). To determine if it has SIFp, though, we need to know if the person fell backward, if they fell farther than 4 feet and where they fell. In our review of thousands upon thousands of incident narratives, it’s rare that both the exposure and essential details are captured.

Jargon also problematic

Lacking details in the case narrative is one problem we often come across. Another is industry-specific jargon and spelling errors. For example:

1. “Can dropped in aisle 45 B, nearly striking employee.”
2. “XD-99 struck XD-87, causing employee to bump their head.”
3. “Employee broke ankle when jumping from truck.”
4. “Bromidine spilt and employee inhailed some. Was taken to nurse’s station.”

Each incident had SIFp:

1. A “can” for this organization is shorthand for a 40-foot cargo container.
2. The XD numbers represent the company’s designation for their forklifts.
3. The fall was from a large piece of mining equipment. The brakes failed when the truck went downhill, and the employee decided the best path of action was to jump.
4. The actual chemical was bromine, and it was inhaled.

Apart from spelling and grammar corrections, the case narratives can’t be adapted at the stage of learning.

Transformer models are a great technology advancement, as embedded words can be adjusted based on context.

However, to achieve a level of understanding in the algorithm, large amounts of data are required.

What have we learned?

Developing an algorithm that can predict SIFp is possible for an organization that has thousands of quality case narratives and common terminology. In both cases, the algorithm can be trained to identify the exposure that contributed to the incident and SIFp.

However, to benefit from an objective, reproducible and instantaneous SIFp classification, more organizations need to input essential information into the reporting system to create more data.

At DEKRA, we have developed an expert system methodology that allows for immediate classification of SIFp, making the decision tree methodology obsolete and is a first step toward moving toward using AI. In our September column, we will explore this topic further.

Direct to your inbox: Sign up to be notified in email about new "Safety Leadership" columns.