Using predictive analytics can arm OSH professionals with a powerful tool to expose critical risks and, potentially, avert future fatalities and injuries.
As Matt Clay, principal engineer at the GB Health and Safety Executive’s Science and Research Centre, pointed out in an IOSH presentation in July 2019: ‘Safety performance has plateaued in many established economies and major accidents continue to occur.’
Coincidentally, a wealth of valuable OSH data is being harvested at a rapid pace. Although this ‘big data’ has the potential to reduce the likelihood of such accidents, Matt says it is often ‘chaotic, unstructured and not necessarily created with “users” in mind’. This is where predictive analytics could be a game changer.
Drawing on statistics and modelling techniques that use artificial intelligence (AI) and machine learning (Suman, 2021), predictive analytics takes historical data captured in incident management systems, such as near misses, and uses it to raise issues of concern.
‘Where predictive analytics gets interesting is where you start to look at the combinations of different sets of data and how it can inform the future of how safety and risk is measured,’ explains Rob Davis, vice-president of product management at information services firm Wolters Kluwer Enablon in North America.
‘We often think about incidents as the basic piece of information that we capture and predictive analytics would be the location, so, for example, what did the incident look like? Who was involved? What were their injuries? Was the right personal protective equipment used? However, other things contribute: what was the weather outside at the time? Was there an anomaly going on in the pressure or something in one of the assets or pieces of equipment that people were maintaining?’
Act before event
By mining historical data in the context of this broader environment to train environment, health and safety (EHS) software’s algorithms to identify patterns and trends, predictive analytics can monitor events in real time and stop potential incidents before they happen.
‘To feed predictive analytics, clients are leveraging technology to increase near misses or unsafe behaviour, for example, automatically creating events by scanning CCTV live feeds based on exclusion zones,’ notes Rob Leech, principal product development director at EHS software firm EcoOnline Global in the UK.
‘By creating frictionless or automatic reporting you are building a picture that is easier than someone typing data into a computer all the time. It’s creating a large dataset. One of predictive analytics’ goals is to highlight things you didn’t know to ask for or were able to find due to the amount of data. The best bit is the insights; the next stage where it is telling you, “I know who you are, where you are and you should know about this to keep you safe.” That’s what we are really focusing on; the insights on the back of those predictions.’
Strategically, what this means is that OSH professionals can be more proactive, flagging up gaps in their systems and processes. Importantly, it also improves resource allocation.
Predictive analytics uses a number of machine-learning techniques to build a more detailed picture of the OSH landscape, including decision trees and Poisson models (Ajayi et al, 2018), and there are three common forecasting models for managing OSH risks – see Predictive analytics models, below.
However, as much as predictive analytics is a great enabler, safety management software also presents challenges, notably encouraging user adoption and ensuring data is recorded on an ongoing basis.
Everyone interviewed for this feature emphasised the importance of EHS software providers consulting OSH professionals early on before predictive analytics technology is introduced and ensuring these subject experts continue to have a hands-on role. ‘It is up to the EHS software provider to ensure the system is user-friendly so the adoption of the technology is high,’ says Dan Hobbs, CEO at EHS software company Protex AI in Ireland. ‘It should be easy to use, fit into their existing OSH systems and mould around their specific problems.’
Another significant challenge is the fact that, currently, there is no standardised data terminology which is accepted by all industries.
In the UK nuclear sector, the Operational Experience Learning Group has created event categories to which sectoral organisations can map their incidents, enabling meaningful comparisons to be made, but this is not a common approach.
A broad adoption of standardised data terminology across all industries would be advantageous, driving greater consistency in data reporting.
Axel Elvik, vice-president, product, at EcoOnline Global in Norway, says the Nordic EHS software provider partners with an aviation project where group members have an agreed nomenclature and have created centralised mapping so they can benchmark against each other’s OSH performance.
He argues that governments and governmental bodies have a role to play in defining standards and encouraging industries to cooperate and standardise different data structures.
Marko Vuorinen, principal business analyst for EcoOnline Global in Finland, adds that this Nordic nation has successfully standardised observations and inspection methods for different industries.
‘When safety inspectors visit companies, they use a scientifically validated method with standardised questions and accepted criteria, where you can see if something is in order or not,’ he explains.
‘The result is a safety index that reliably and proactively indicates the safety level of the company or area, even if there have been no accidents. By combining huge observational data with accident data at the national level, it has been possible to determine the correlation between subfactors, like order of walkways and injury risk. Then even individual companies can trust the method and use it to continuously improve their safety level.’
As Tjerk de Greef, director of product software engineering, Wolters Kluwer DXG in the Netherlands, notes, data cleanliness is critical to any good predictive analytics model.
‘The models can only be as good as the data we train the algorithms on, so if the incidents are not reported, we can never predict them and you will always need to clean up the dataset.’
The AI system does all the predictive work, identifies if there is a big spike in near misses and gives that data to the OSH professional
Early warnings: Predictive analytics models
HSE Network identifies three different predictive analytics models used in safety.
Forecast model: This centres on ‘specific metric value prediction’ with the model learning from historical data and generating new forecasts. OSH teams that use historical data can apply this model to any safety situation. One example given is a safety manager who wants to predict the number of missed or faulty risk assessments that could happen over a specified period.
Classification model: This involves classifying data generated from inputs into the model. One example given is to determine the likelihood that a specific piece of equipment is faulty in a certain timeframe.
Outliers’ model: This approach can identify ‘anomalous data within a series’. The model categorises data but if the inputs aren’t correct, it can identify anomalies so OSH professionals can resolve any safety concerns raised.
Limits of prediction
Dan points to a predictive analytics tool that uses AI-driven camera software installed in warehouses and ports. The technology, which collects data anonymously, highlights areas of concern for OSH teams to act on early. ‘The AI system is doing all the predictive work for you and can identify if there is a big spike in near misses and gives that data to the OSH professional so they can make the decisions.’
But what can’t predictive analytics do? Rob Leech says the technology will struggle to predict all accidents: ‘How can technology know that Bill the FLT driver is in a bad mood which is affecting his attention today?’
The human element, such as good reporting and investigations, remains essential, with AI as a complementary tool.
Rob Davis says predictive analytics isn’t always about identifying future potential risks; the technology can also remove some of the laborious and time-consuming manual work that managers do, freeing them up to dig deeper into data analysis.
‘You don’t know what you don’t know,’ Rob Leech adds. ‘If you’re looking for a risk, it’s because you already know it’s there, and analytics can drill down to the cause and fix it. But when you don’t know problems, analytics can tell you very quickly too – it can watch the anomaly detection and say: “These are dropping off quicker than usual: is there a problem?”
‘It provides managers with real-time, up-to-date analytics, with a big enough picture to tell them what they didn’t know.’
This proves invaluable when companies are too busy logging data and looking at incidents that have happened to notice near misses, agrees Rob Davis.
‘Having those near misses builds better predictive analytics. We are working with a major high-tech manufacturer that uses our technology to take incidents and lump them into categories using cluster analysis. The system says, “All these incidents look similar. I can assign the same mitigation to it and get them off my incident manager’s plate”. This gives the manager more time to look at near misses, and it makes the algorithms better.’
When implementing predictive analytics technology, Axel says it is important that organisations consider a few critical points.
‘Be very clear on the goals and set a few key metrics that you choose to track, and then you can get the company to rally around because ultimately it is about driving awareness and change,’ he says. ‘Communicate those metrics and visualise them all the way down to the frontline workers. You’ve got to get into this continuous improvement mindset.’
Rob Davis says that when organisations first start using the models, they will rely on the algorithms provided by their EHS software provider. However, over time, they’ll want to merge these models with other algorithms and capture external data.
‘As companies mature, they will be employing data scientists,’ he explains. ‘They will be employing analytics specialists to make sure that that data gets pulled in. This is where the openness of the solution is important, so you can integrate it into your third-party analytics.’
And while smaller organisations can benefit from AI too, there are limitations, says Rob Leech. ‘The bigger the dataset, the better it is, the more reliable it is and the more capability you have,’ he explains. ‘Unless it's built in to the product they’re buying – whether it’s anomaly detection, or notifications based on an analytic – there’s not a lot of great leaps for AI aimed at small datasets and therefore small businesses. That said, the models are getting better all the time.’
Data governance and privacy
But what about sensitive issues, such as personally identifiable information? Tjerk says it is critical that OSH professionals consider data governance requirements carefully, particularly where data can and can’t be stored, and privacy issues.
‘We capture information from particular people,’ he explains. ‘You are not interested in the person themself but in the employee ID, so you want to be very careful that not everybody can access this data. You need to govern that properly; so data encryption is important, and you must respect all privacy concerns. There are regulations, but there are also ethical guidelines you will want to follow and implement. If you take videos, for example, you can blur people’s faces.’
Predictive analysis that uses big data is still in its early stages; so what is the future of this technology? Rob Davis argues it comes back to having an open strategy that enables data to be pulled in from third parties. For example, Enablon’s software records the performance of contractors on oil and gas sites, including accident records and hiring practices.
‘All of these things can be captured in the software and are increasingly going into the environment, safety and governance metrics that companies are reporting. These are things that can be used to say: “What is the risk to a company of using a given contractor?”’
Grasp the potential
Dan says the future of predictive analytics is already here. One of the positive developments is that OSH professionals are much more tech savvy than they were in the past; they grasp the technology’s potential and want to use it.
For Rob Leech, the bigger picture, and a much more long-term development, is the potential for growth in the machine-learning services and consumable predictive analytics specific to particular sectors.
‘Predictive analytics and machine learning go hand-in-hand, and over time the models will train and grow based on sector specific large datasets providing tailored and focused predictive analytics options.
‘I feel we are still on the hype curve but I do see a future with more ready-made trained models for your industry to try and help you with predictive analytics.’
- Role of data and horizon scanning in management of change: iosh.com/media/5146/mc-presentation.pdf
- Is EHS software the Holy Grail? ioshmagazine.com/2022/06/28/ehs-software-holy-grail
- AI in OSH: a smart move: ioshmagazine.com/2021/02/19/ai-osh-smart-move
Ajay I, Oyedele L, Delgado MD et al. (2018) Big data platform for health and safety accident prediction. World Journal of Science Technology and Sustainable Development 16(1).
Suman A. (2021) Rethinking predictive analysis: Learn how to stop workplace incidents before they occur. ISHN. (accessed 5 September 2022).