How does OHS Get the Most out of AI and Machine Learning?

There are two critical questions OHS professionals need to ask in order to design the right approach for applying artificial intelligence (AI) and machine learning in order to deliver optimal functional and business outcomes, according to an expert in the area.

OHS professionals can play an empowering role in this process, as data scientists generally don’t have the subject matter expertise required to understand what problems need to be solved in an organisation.

Instead, OHS professionals are the best people to provide answers to this, said Lok Yiu, senior WHS insights and analytics officer at the NSW Government’s Centre of Work Health and Safety.

“In my experience, prior to model building, 70 per cent of my time was spent in figuring out only two questions,” he said.

The first one is “what is the biggest ‘problem’ worth solving in your organisation?”

“For example, a defined problem can be: out of your 25 warehouses, which warehouses will have at least one serious musculoskeletal injury that costs more than $10,000 in workers’ compensation, happening in the next 365 days?”

Yiu, who was speaking ahead of the SIA National Health and Safety Conference, which will be held from 22-23 May 2019 at the International Convention Centre in Sydney, said OHS professionals and subject matter experts play a key role in understanding what OHS problems really matter and getting involved in translating a business problem to a data problem.

“This process is critical to enable data scientists to draw assumptions and develop an algorithm that align with the business problem,” said Yiu.

The second question OHS professionals need to ask is whether existing data is sufficient to support a reasonably confident answer for the problem(s).

Existing data can include workers’ compensation data, inspection records, leave history and workers’ average ages, for example, and Yiu said data scientists need to work collaboratively with OHS professionals to help understand this data.

“In most cases I have spent a substantial amount of time with inspectors to understand the reliability of data and the data workflows – for example, how the data is being entered, existing data validation rules, and potential sources of errors,” he said.

“OHS professionals play a key role here to fully release the ‘potential’ of your data by adding subject matter understandings.”

As a data scientist, Yiu said OHS professionals don’t need to understand all the algebras or the arithmetic that sits behind machine learning models (this is a data scientist’s job) and he observed that the question of how OHS professionals can add value to the process is not well-discussed.

“Speaking as an OHS professional, we are probably looking for AI to power a lot of solutions that can reduce costs, improve accuracy and augment risk management decision making,” he said.

“To archive these goals, we need to firstly solve the two fundamental questions I mentioned before.

“Following that, the next question would be: should the technical part of the problem be solved by an in-house team or be outsourced to a third-party data science service provider? “Both options have their own pros and cons and need to be fully justified,” he said.

Assuming a decision has been made regarding this question, Yiu said the final issue for OHS is how well project outcomes are sold internally in an organisation.

“This is a process of translating a data answer back to a business answer,” he said.

“This is something that an OHS professional could do to provide a strong influence in deciding how to control the risk and educating people,” he said.

“A machine learning model will only be able to tell you how likely your warehouse will have a serious workplace injury; we as OHS professionals need to give answer on what is the best controlling strategy and then, selling this to the management and to staff,” he said.

Yiu also observed that there will be “plenty of opportunities” in applying AI to OHS in the future.

Currently, one of the key roles for OHS professionals is interpreting data from various sources, usually in the form of an integrated key performance indicator (KPI) dashboard that captures historical WHS performance, such as the number of injuries or incidents by months, or the cost of workers’ compensation claims.

Yiu added that a major limitation of this current approach is that sometimes KPIs are conflicting, and the KPI itself – usually a number or a percentage – does not provide actionable insights.

“For example, we see the past quarter’s average cost per claim has increased by 9.0 per cent; is this good or bad? How can I tell if I should take actions or not?

“In addition, the historical KPI data (usually focusing heavily on ‘lagging’ indicators) only confirms our knowledge of what has happened.

“If top management relies purely on the dashboard in making risk control decisions, such an approach seems to be reactive, delayed and probably biased,” he said.

In terms of future trends, Yiu predicted one of the biggest breakthroughs in the next five years would be the use of large-scale, real-time and automated AI methods to flag OHS related-anomalies in day-to-day operations – and using this real-time data to prevent incidents and injury.

“A perfect example would be the ‘Google flu trends’ (GFT) services, which predicts and tracks a flu spread, based on a large number of real-time Google search queries about flu symptoms,” he said.

“It was reported that GFT was able to predict flu outbreaks up to 10 days before they were reported by the CDC (Center for Disease Control and Prevention).

“In theory, if we can identify or enable a powerful real-time predictor, we can also use a similar approach to learn about workplace injuries and illnesses,” said Yiu.

Article originally published by the Safety Institute of Australia.