TabPFN helps clinicians predict intubation - before it’s too late

Researchers at the University of Warwick, in collaboration with the NHS, are piloting a new application powered by TabPFN to support decision-making in the treatment of Acute Respiratory Failure (ARF). The project is currently in real-world pilot testing at a major UK non-invasive ventilation (NIV) centre, led by Surgeon Commander Tim Scott at the University Hospitals of North Midlands NHS Trust.

The Clinical Challenge

Acute respiratory failure (ARF) is a common condition in emergency and intensive care units. Clinicians often begin with non-invasive ventilation (NIV) to avoid the risks of invasive ventilation, but this approach fails in up to 40% of cases. Delayed intubation after NIV failure significantly increases mortality risk, making early prediction important for patient outcomes.

Current decisions rely on clinical intuition, guidelines, and scoring systems. Existing clinical scores such as HACOR or ROX require serial assessments and have limited generalizability across patient populations and care settings. As highlighted in recent studies (Vargas & Servillo, 2022, Hu et al., 2022), no single physiological parameter or simple scoring system has proven sufficiently accurate to guide this decision in isolation.

The Solution

The pilot uses a web-based application integrated into the NIV workflow:

For NHS clinicians:

• Patient data is entered at the start of NIV treatment and again at the 2-hour mark
• Predictions are generated using TabPFN, running on GroqCloud infrastructure
• To avoid influencing clinical decisions during the pilot, predictions are reviewed only by researchers, not treating doctors

For researchers:

• Receive patient data from the NHS centre
• Run predictions through TabPFN on secure servers
• Compare predicted outcomes with actual results

Technical Advantages

TabPFN's performance in this pilot demonstrates several key capabilities:

• Small dataset performance: Effective results with limited medical data
• No preprocessing required: Works directly with raw clinical measurements
• Rapid deployment: Train and evaluate within seconds
• Real-world integration: Can be deployed in an app for bedside use

Impact so far

In the current NHS pilot testing, TabPFN correctly predicted 22 out of 24 patient outcomes, demonstrating potential for AI-assisted decision-making in high-pressure, time-sensitive environments.

Surgeon Commander Tim Scott, Consultant Anaesthetist 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.”

Next Steps

Building on the promising results from the initial pilot, the research team is preparing for the next phase of validation. They aim to launch a randomized controlled trial across multiple NHS NIV centres to evaluate TabPFN's deployment at scale. This expanded study will provide crucial evidence for the model's effectiveness across diverse clinical settings and patient populations, bringing the technology closer to widespread NHS implementation.

The study was published in the Intensive Care Medicine Journal and featured in Medical Xpress.

‍