Data analytics & AI - The Future of OHS
Shirley Parsons sat down with James Pomeroy, Group Health, Safety, Environment and Security Director at Lloyds Register, to discuss his opinion on the future of OHS and how Data analytics is set to change the way practitioners view Health & Safety for good.
Why are you investing in data analytics at Lloyd's Register?
While the OHS world has always used data, as a profession we have historically been quite limited in our understanding and application of it. Almost all organisations collect OHS data, such as the number, type and frequency of accidents, injuries and ill health. Not many, however, go beyond this to fully realise the potential of the information at their disposal. Furthermore, while there is a lot of OHS data, most of it is contained in ‘silos’ and not integrated.
In essence, while practitioners have vast amounts of OHS data at their disposal, most of us have not historically utilised it to its full potential. With the advancements in technology, we now have a great opportunity to apply the techniques of data science that have been successfully applied in the marketing, financial, and insurance sectors to OHS.
Can you provide some examples of the potential advantages that you see?
In OHS, we have traditionally focused on lagging indicators, notably the types and numbers of accidents. While we routinely make interventions to address specific hazards, such as changing equipment, conducting training or revising the working methods, the link between these interventions and the impact on outcomes is often weak.
For example, we see the number of injuries declining over a period and assume it is because of our recently introduced programme or changes, but are the measures really correlated? Much of the decision-making in OHS remains intuitive. Data analytics can bring greater scientific rigour and introduce more evidence-based decisions. This could be key when considering different types of intervention or when promoting a strategy that involves considerable cost.
The potential also exists for practitioners to link the different types of OHS data we have, so we can make more informed decisions. Most organisations have a wide range of ‘data sets’, including training and competency records, employee demographical information, and health data, task and project information, alongside audit, inspection, observation and near miss reports. Some of these data sets are measuring safety interventions, others provide an indication of individual outputs, and the best give us an indication of true outcomes, such as whether employees are healthier and safer at work. These data sets are rarely integrated and we often implement OHS strategies and improvement actions in isolation.
A third benefit is being able to address the volume of data that a good OHS programme creates. Most progressive safety cultures aim for high levels of employee engagement, either through safety observations, ideas, or near miss reports. If successful, this can create huge amounts of reports, all of which need to be reviewed and analysed. This is time-consuming and a common failure in safety reporting programmes is the inability to handle the volume of reports that employees create. The introduction of artificial intelligence into incident reporting systems has the potential to simplify the reporting process and automate the analysis, enabling practitioners to utilise their time more effectively.
The potential of this technology is, however, much greater than merely the gathering or analysing of incident reports. Much of our thinking in OHS is driven by theories that are very conceptual and often of questionable scientific value. This is not solely limited to historic models, such as Heinrich’s infamous triangle, but equally applies to accident causation models that have emerged over the last three decades. OHS is full of what Feynman infamously called ‘Cargo cult science’. The application of data analytics will test these models in ways that we were previously unable to do. We’re entering some very interesting times in OHS!
Why has this not been done before?.
Correlating different types of OHS data is not necessarily new and some organisations already do this. What is new is the ability to introduce real-time monitoring data and the introduction of artificial intelligence into OHS. For example, many organisations now deploy telematics that provide real-time data on their driver behaviour; other organisations are utilising wearable technology to monitor and record personal exposure to noise, vibration, heat, etc. These technologies create a vast rich new set of information that can be integrated to other potential sources of information. For example, integrating driving hours and scheduling, with the open source data on traffic and weather could be transformative for managing driver fatigue and preventing micro-sleeps. Add in the outputs from wearable that monitor driver fatigue and you have a vast arsenal of potential information at your disposal to keep people safe. You can apply similar examples to other high-risk activities such as working outside in the heat or at height.
One of the most exciting areas for us at LR is the rapidly evolving area of artificial intelligence – this could be a game changer for how we collect, analyse, and understand OHS data. Most practitioners will know how challenging it can be to collect data from personnel on their OHS experiences, particularly near misses, observations, and ideas. But these reports take time for personnel to record and analyse. It's also necessary to recognise an inherent contradiction in the process to collect such reports - the more information we ask for, the less we actually get. This is because many OHS forms are time-consuming to complete and the really useful information is contained in the description of the incident.
AI and the use of voice recognition to automatically populate incident reports will make it significantly easier for employees to report their experiences, letting the technology do the work and categorise the incident. Even more exciting is the ability for AI to ‘self-learn' in how it analyses data and this only now being realised. Incident reporting and the provision of OHS management information will certainly look a lot different in 10 years as a result of AI.
There are also contextual reasons that this technology is now being applied to OHS. The costs of safety failures are now significant for organisations in financial, operational and reputational terms, and this makes it easier to justify the investments required. It is also worth noting that most organisations have tried all the traditional OHS interventions and are looking to try more evidence-based interventions.
Can you provide some examples of where this is already helping you improve performance?
Like many organisations working in the high-risk sectors, at LR we place a high value on employee engagement and encourage our personnel to share their experiences through near miss and safety observations. We’ve had some great success here and regularly receive over 700-800 reports each month, each containing 16 separate items of data. Understanding what these reports are telling us, and what trends we can ascertain, is a huge challenge for us that is time-consuming. It’s somewhat ironic that the better we are at getting employees to share their experiences, the more of a challenge we have in reviewing, categorising, and responding to them. Working with our data science team, we have applied AI to interpret thousands of employee descriptions to highlight specific trends that are not present in the 16 categories we ask employees to report upon. In the case of driving, for example, applying AI to the thousands of reports tells us that the distance, stop light, and seat belt are words frequently being reported. This is helping shape our thinking about road safety. We have similar lessons from looking at our other significant risks, such as confined space work and pressure testing.
Another real-life example for us is sharpening up the taxonomy of incident reports. When reporting an incident, we ask employees to categorise the type of incident. The categorisation helps us analyse the incident and run trends, but approximately 20% of employees chose the "other" category when reporting, making it difficult to understand what they’re reporting. Analysing these 150 reports per month is very time-consuming. When we apply an algorithm to this category, we can identify what the incident is really about by the keywords used and then correctly allocate it. The really exciting thing is that the algorithm self-learns and thus improves its accuracy over time.
What advice would you share with practitioners interested in this subject?
Almost all organisations will have lots of data and it could be easy to get lost in the maze of reports and data that you will likely encounter. At LR we started by identifying and mapping all of our data sets and, with the help of our data scientists, classified each into leading and lagging sources of potential data that we could correlate.
Once they look, most practitioners will discover lots of potential data sets and it could be quite easy to develop lots of analytical studies with no clear outcome. For these reasons, I would therefore suggest starting small by selecting one specific problem or opportunity and focusing on tackling this first.
It can be a complex area and unless you have good statistical skills, I’d suggest practitioners get a partner to help them. Most practitioners will not have the technical skills and there is some real value in utilising someone who is not an OHS practitioner, as they look at issues with no preconceptions and are comfortable challenging the current thinking.
Almost all organisations will have issues with data quality. Most studies will involve some element of data cleansing and this needs to be considered from the outset. As previously highlighted, it’s also essential to understand the regulatory and ethical issues, particularly when handling employee data, injury reports, and information on wearables. Some of these data sets will be held by other functions and they, understandably, could be cautious about sharing this information.
From a personal level, I would also encourage practitioners to ‘be curious’ about this and expect to have their thinking challenged. OHS is a very traditional field that for decades has had blind faith in theoretical models that were rarely questioned. As a profession, we also tend to act in herds and think collectively, often applying the same interventions as our colleagues because “it’s worked there and thus, it’ll work here”. Data science will provide evidence-based arguments that allow much more targeted and appropriate interventions.
What are some of some of the challenges you foresee?
Some of the OHS data is personal to individuals, highly sensitive, and governed by the new GDPR regulations, notably health and medical data. We therefore need to make sure that we fully explore the regulatory and ethical issues from the outset. The emerging field of data ethics presents some real challenges for us. Some of these issues can be overcome by looking at the data at a macro-level to identify trends.
Aside from the regulatory and ethical challenges, there is the obvious issue of the analytical skills to undertake this work. Data science and advanced statistical analysis do not currently form a significant part of the OHS training, either at diploma or degree level. As a minimum, practitioners need to have an idea about the data sets they hold, where data science could be deployed in their OHS programme, and how to interpret the information that they're provided by the experts. For this to happen, the professional will need to introduce a greater focus on training and professional development. It's interesting to note at LR that we're working with a team of data analysts that have not previously worked in the OHS arena before. Fortunately, OHS data is relatively straightforward and our operating models are simple compared to other business disciplines they encounter.
The quality of OHS data can often be a real challenge, particularly historical data. There are also very basic issues such as the format of the different records, the language that the reports are submitted within (particularly an issue for global organisations), and the incompleteness of many records. Time is required to evaluate and ‘clean the data’.
Finally, I think there is a challenge around managing expectations. There is lots of talk of predictive analytics – the ability to predict accidents before they occur from historical data - but that remains highly problematic. There is very rarely one cause of an accident and the multifactorial nature of most incidents means that predictive analytics currently remains the stuff of movies like Minority Report.
A lot of the application of data science has also been in closed systems and every practitioner will tell you how complex human behaviour can be. There are, however, some very clever people looking into this area now and, with the fast-moving area of artificial intelligence, who knows what tomorrow brings?
So you see this as the future for OHS?
We commonly hear the phrase disruptive technology, but the application of data science in OHS truly has the potential to transform how we think and work as practitioners. These are really exciting times to be working in OHS. My sense is that the new thinking and new technologies being developed now will make our jobs very different in the years to come.
About the Author:
Shirley Parsons are global leaders in HSEQ Recruitment, search and staffing services. We are an ever-growing global HSEQ talent network built on long-term relationships, industry knowledge, and geographic expansion.
View our latest jobs or get in touch.