A New Zealand Trial Asks Whether Machine Learning Can Personalise Life Support

Most stories about artificial intelligence in medicine describe a tool that is already running. This one is about a question that has not been answered yet, and about the deliberate, expensive machinery a New Zealand research team is building to answer it. The question is narrow and very high stakes: when a patient is on life support in intensive care, can a machine learning model help doctors decide how much oxygen that particular person should receive.

The trial is called the REVOLUTION trial, and it is led by Professor Paul Young, Deputy Director of the Medical Research Institute of New Zealand and co-clinical leader of the intensive care unit at Wellington Hospital. It has been funded with a $5 million Programme Grant from the Health Research Council of New Zealand, one of two Programme Grants and 38 Project Grants, worth close to $57 million in total, announced by the HRC on 11 June. According to the announcement, it will be the first major clinical trial in the world to evaluate whether AI-guided treatment using machine learning improves survival in the ICU.

Oxygen is a deliberate choice of target. It is the most widely used therapy in intensive care, and about twenty million people worldwide need life support each year. It is also a therapy where “more” is not obviously “better”, and where the right level may differ from one patient to the next. That is the gap the trial is built around. As Professor Young put it, “It is easy to imagine a scenario where a treatment benefits a population ‘on average’ but where only a small group of patients benefit greatly, and most are in fact harmed. Machine learning could provide the information clinicians need to ensure they give the treatment only to those patients who will benefit and not to those who will be harmed.”

The design is worth understanding, because it is where the credibility sits. The team will first refine a machine learning model using data from Mega-ROX, a recently completed trial that Professor Young’s group describes as the world’s largest ICU trial, involving 40,003 patients across 137 ICUs in 14 countries. The model will combine that data with information clinicians enter when a patient is admitted, and estimate the individual benefit or harm of higher versus lower oxygen levels for that person. REVOLUTION will then compare two approaches head to head: oxygen delivery guided by the model, against the standard approach where doctors decide without it. The trial is planned to run across 50 ICUs in New Zealand and Australia and to recruit more than 24,000 patients, with the New Zealand component funded by the HRC and the Australian component by the NHMRC.

The significance is in the method as much as the medicine. Personalised, or precision, care has been promised in many corners of health for years, and intensive care is one of the hardest places to deliver it, because decisions are made quickly and the patients are fragile. A randomised trial that puts a machine learning model to that test, rather than assuming it helps, is the kind of evidence the field has been short of. Professor Young is careful about the size of the prize, framing it in terms of a “modest reduction in deaths” that, spread across the millions of people who need life support, could still mean hundreds of thousands of lives each year.

None of that is proven yet. REVOLUTION is a trial that has been funded and designed, not a result that has been delivered, and its value will be decided by what the data shows once every major New Zealand ICU and dozens of Australian ones have taken part. What it does offer now is a clear, checkable example of AI being introduced into critical care the slow way: through a large, randomised, clinician-led study rather than a product launch. The full announcement, including the other funded research, is set out in the Health Research Council’s grant round.