Prioritising referrals by need
Heart disease is the leading cause of preventable mortality in Aotearoa New Zealand. It’s also a condition that disproportionately affects certain groups, such as Māori and Pacific peoples.
In order to prevent serious illness early medical intervention is essential, but this is dependent on patients having access to medical care. This is not always the case for Māori, whom data shows to have higher rates of coronary artery (heart) disease and stroke than non-Māori, but lower access to health care.
Kevin Ross, CEO of Precision Driven Health (PDH) – a collaboration between Orion Health, Te Whatu Ora – Waitematā and the University of Auckland aimed at improving health outcomes through data science. – says: “What we know overall is that Māori have worse outcomes than Pakeha in the health system.”
“You’d assume, for example, that because Māori have twice the rate of coronary disease, they would also have twice the referral rates, but we don’t see these referrals coming in at quite that same rate.”
The triage process may see some patients with lower need being accepted for a hospital clinic appointment, moving people who are more in need further down the queue.
There is also a growing backlog of referrals too, which – if a clinician works in the order they are received, as is the case currently – could leave higher-need patients in the backlog awaiting triage.
Investigating inequities has been part of a PDH-supported research project undertaken by Orion Health which developed a deep learning tool to help improve the triage process from GP referrals, to improve health outcomes for patients through efficient and timely processing of their referral
Analysing the triage process
Deep learning is a type of artificial intelligence (AI) that imitates the way humans gain knowledge. It’s an important element of data science, and makes the task of collecting, analysing and interpreting large amounts of data faster and easier.
Little research had been done to improve the triaging process in New Zealand using deep and machine learning techniques. This is partly due to the challenging nature of working with electronic GP referrals in New Zealand, which contain both structured and unstructured (free text) data.
To help address this, PDH put deep learning to use to analyse the process of triaging GP referrals to cardiologists – a time-consuming task for doctors in New Zealand, who may commonly spend more than 10 hours per week on triaging referrals to various risk categories or ‘direct to investigation’ tests.
It was a switch from paper-based referrals to electronic referrals, however, and analysing the data these capture, that enabled PDH researchers to gain a better understanding of who is being prioritised for early referral to cardiologists.
A key objective of this project was to identify potential bias in the existing triage dataset, and provide a machine learning solution that mitigates implicit bias, reducing the issue of access by moving patients most in need up the queue.
The research team analysed available anonymised patient records from the Te Whatu Ora – Waitematā, with referrals to cardiology specialties from 2015 to 2018 included in this dataset.
In the analysis, there was no observed difference in triaging between Māori and non-Māori once patients have been referred for an appointment. However, an argument can be made that a higher priority to Māori would help to counter the known differences in overall outcomes. This would require preferential bias/affirmative action to be applied.
PDH’s research provided the opportunity to identify and address bias, by first acknowledging and identifying biases and inequity and then testing ways to change these with Māori involvement.
To do this, PDH worked collaboratively with Māori health researchers to develop strategies to identify and address any implicit bias, in terms of unequal access to specialists arising from GP referral triaging.
The focus was on trying to counter bias to provide an equitable solution – one that recognises that each person has different circumstances, and allocates the resources and opportunities needed to reach an equitable outcome. Equity needs to be addressed at every step of a health journey and triaging is only a small part of this.
Collaboration is key
Haze White, a Māori health advisor who has played a key role in this project, says: “For data science to be effective the data it uses needs to accurately represent the problem at hand. For example, the people who could really benefit might not be those who go to GPs, or alternatively, go straight to the hospital.”
“When we spoke to GPs and some of the specialists, ethnicity was not formally used as a variable to triage a patient’s priority. That can be something that the GP referrals triage tool does.”
Haze worked with Dr Ning Hua, a Senior Data Scientist from Orion Health, to extend the data set used in the GP referrals triage project, with Ning saying that their collaboration helped to create positive outcomes.
“Initially when we looked at the referrals dataset, we didn’t observe any inequity among different ethnicity groups. Everybody who was referred was being prioritised quite proportionally,” says Ning.
Data science provides an opportunity to advance the equity discussion. Data scientists and domain experts can collaborate to better shape what fair outcomes look like, and tune the modelling towards this. If we don’t do this we run the risk of building biases into our models.
Ning Hua, Senior Data Scientist, Orion Health
Haze adds: “We got buy-in from our specialist team at Te Whatu Ora – Waitematā that we could potentially do something different, by applying ‘weighting’ to contribute to your prioritisation based on certain factors like age and ethnicity.”
This research is ambitious in that it aims to change the way data is used to support the health of New Zealanders who need it most. The automated triaging tool that emerges from this research could also be applied across New Zealand as well as overseas.
Through analysing the data, the aim is to create an effective, automated system that will correctly prioritise patients without bias. The ultimate result will be that the most at-risk patients will be prioritised.