New information about controversial ACC prediction tool released

Privacy Commissioner John Edwards says while he wouldn't automatically be concerned about ACC's use of an algorithm, it should be subject to a high level of transparency. Photo / Mark Mitchell

ACC has released further details about a controversial risk prediction tool it uses to profile and target clients, following concerns about privacy, ethics and transparency.

A Herald investigation revealed yesterday the corporation had used historic claims decision data to build the model behind closed doors in 2014, without client consent or knowledge.

It predicts how long clients are likely to be on ACC's books, information which is then used by staff to undertake "proactive management" if people aren't meeting expected recovery times.

Advocates are concerned about potential discrimination arising from the use of the tool. Overseas such models have been found to be biased and inaccurate.

They were also worried about its potential to drive unethical behaviour - for example pushing people back to work when they weren't ready - given ACC's performance was measured by duration-driven targets.

ACC minister Michael Woodhouse said those concerns were unwarranted, and to suggest the model limited people's entitlements was wrong.

However, he revealed he did not know exactly how the model worked - even disputing it was predictive - or whether it had been through ethical review.

Last night, ACC released a statement on its website with more information about the tool, which it said was built from 364,000 claims that were lodged between April 2007 and May 2013.

It said clients signed a privacy notice about data collection. The form notes a claims database but does not say that ACC uses predictive modelling.

Factors used in the model included diagnosis, age, lodgement delay, work type, questions about whether treatment was sought at a DHB, a client's injury history and initial incapacity as outlined on the claim form. It did not say if that was a full list of factors.

The model was last reviewed externally in 2014 and updated in 2016 and 2017, it said. Details of reviews were not provided.

"This is work done solely for ACC: other agencies do not have access to the data," ACC said.

It said the tool's benefit was in helping ACC engage quickly with clients. Since it was introduced, the average time taken to set up weekly compensation for clients has dropped from 11 days to seven days.

It had also seen an increase in customer satisfaction.

"Case owners are now able to make evidence-based decisions - something we think is important," it said.

Yesterday, opposition parties said people had the right to know what the government was doing with their data, and how decisions were made about them. Both Labour and the Green Party said the Privacy Commissioner should have been involved in the model's creation.

Advocates Acclaim Otago said they believed it added weight to calls for a commissioner to oversee ACC.

Privacy Commissioner John Edwards said while he wouldn't automatically be concerned about the use of an algorithm in ACC's case, it should be subject to a high level of transparency, and not overly relied upon as a substitute for human decision-making.