Researchers at the University of Warwick have developed an AI tool to support doctors in the complex decision of whether to intubate a patient with acute respiratory failure. The tool, based on standard clinical data such as respiratory rate and oxygen levels, predicts within two hours of starting non-invasive ventilation (NIV) with above-average accuracy whether this treatment will fail.
Although NIV is often the first treatment option, this method fails in approximately four out of ten (40%) patients. Moreover, it sometimes leads to serious complications. Deciding to switch to invasive mechanical ventilation requires quick insight, while doctors often operate under time pressure with only limited data. According to Prof. Declan Bates, principal investigator and professor of Engineering Sciences, the new AI tool offers a solution by providing data-driven insights that would otherwise go unnoticed.
Pilot study
TabPFN is specially designed for tabular patient data and uses “in-context learning” to make predictions without extensive training data. In a pilot study conducted at University Hospitals North Midlands NHS Trust, the model is being tested via an app. Doctors enter routine NIV measurements and receive real-time feedback on the likelihood of treatment success. These predictions are later compared with actual patient outcomes to further confirm their reliability.
‘The accuracy of the AI tool is impressive. We can clearly see how early prediction helps improve treatment decisions and patient experiences. If this can be rolled out on a large scale within the NHS, it will be a valuable innovation tool,’ said Tim Scott, an anaesthesiologist at the hospital.
Support, not replacement
Like all AI tools and applications, TabPFN is not intended to replace doctors, but to support them in complex decisions. According to Bates, the model helps doctors by objectively analysing which patients are likely to fail NIV within two hours, allowing them to anticipate this in good time. The final decision rests with the doctor or healthcare professional.
‘Acute medical emergencies, such as respiratory failure, are associated with high mortality rates and resource-intensive care. AI can help treat these patients more effectively and improve care outcomes,’ says Professor Gavin Perkins, Dean of Warwick Medical School.
Earlier this year, we wrote about a new Explainable AI tool that predicts the length of stay for ICU patients while providing a transparent explanation of the factors influencing the prediction. The model, developed by Prof. Indranil Bardhan, analyses 47 patient characteristics at admission and converts them into clear graphs that visualise the likelihood of discharge within seven days and the most important influencing factors. In a study involving six ICU doctors, four doctors found the model's explanation valuable for bed management and staff planning. The system offers similar accuracy to existing models, but with greater insight.
