New AI Tool to support decisions around patient intubation

University of Warwick researchers have led the development of a new AI tool that can help doctors make the difficult and high-stakes decision of whether to intubate a patient in acute respiratory failure.

Acute respiratory failure occurs when the respiratory system cannot provide oxygen to, and/or remove carbon dioxide from, the body. Treatment is primarily based on providing external respiratory support, such as noninvasive ventilation (NIV) through a facemask, but around 40% of patients fail NIV and subsequently require endotracheal intubation and invasive mechanical ventilation.

Published in Intensive Care Medicine, University of Warwick researchers have developed a new AI model that can help clinicians identify patients who will need intubation much earlier in their treatment, which improves outcomes for those patients with acute respiratory failure.

Professor Declan Bates, School of Engineering, University of Warwick, led the study and said, “Each step of treating acute respiratory failure requires clinicians to make critical decisions, in a time-pressured environment, without having access to all the information. Furthermore, patients that fail noninvasive ventilation subsequently have an increased risk of mortality, so these decisions have real consequences.

“We created this AI model to work with the measurements that a clinician would normally make, such as respiratory rate and arterial oxygen levels, and produce a prediction of NIV failure based solely on that data within two hours of starting NIV. It’s significantly more accurate than existing methods, which makes it really promising for testing in clinical trials and ultimately widespread adoption.

“It’s important to point out that this AI model is not designed to replace the decision-making of doctors. Its aim is to assist them by making the best use of the patient data—the AI crunches the numbers in an objective manner to make predictions that clinicians can then factor into their extremely complex decision making.”

The AI model, called TabPFN, is a novel machine-learning model specifically designed for tabular data classification tasks. It uses ‘in-context learning,” so that it doesn’t need to be trained from scratch and can immediately produce accurate predictions when faced with new data (such as small sets of patient measurements).

TabPFN is already seeing its first real-life testing in a pilot study at University Hospitals North Midlands NHS Trust. Using an app version of the AI model developed at Warwick, clinicians input routine measurements from NIV patients at the hospital. This data is fed back to Warwick, where the AI model tells the Warwick team its prediction of whether the patient will succeed or fail on NIV.

Later, the clinicians feedback the actual outcome for the patient, which is compared to the real-time prediction made by the AI model, giving a measure of its accuracy.

Surgeon Commander Tim Scott, Consultant Anesthetist at University Hospitals North Midlands NHS Trust and the Royal Centre for Defence Medicine, Birmingham said, “My colleagues and I are currently testing an app based on this model in our hospital and its accuracy in predicting the outcome of NIV has been extremely impressive. We are very enthusiastic about its potential to improve patient outcomes and hope that further development will enable it to be rolled out across the NHS.”

With no formal guidelines around intubation in place, and evidence that clinicians don’t trust and aren’t relying on the existing measures to help with their decisions, TabPFN is a very welcome development.

Professor Gavin Perkins, Dean of Warwick Medical School, said, “Patients with acute respiratory failure consume a disproportionate amount of hospital resources, mortality rates are high, and survivors report low health-related quality of life. AI has huge potential to help clinicians manage these patients better and improve their outcomes.

“Warwick’s Clinical Trials Unit is at the forefront of evaluating new therapies for acute respiratory failure, and we look forward to working with our colleagues in the School of Engineering to develop these new technologies for the benefit of patients.”