A research team has developed an AI foundation model that analyses clinical notes from surgical patients to predict complications such as pneumonia, blood clots and infections.
The study, published in npj Digital Medicine, suggests that this model could significantly reduce the 10% postoperative complication rate, which contributes to longer ICU stays, increased mortality, and higher costs.
The team is based at Washington University in St. Louis (WashU). Lead researcher Chenyang Lu, PhD, a professor of computer science and engineering at WashU, said: ‘Surgery carries significant risks and costs, yet clinical notes hold a wealth of valuable insights from the surgical team. Our large language model, tailored specifically for surgical notes, enables early and accurate prediction of postoperative complications. By identifying risks proactively, clinicians can intervene sooner, improving patient safety and outcomes.’
Traditional risk models rely on structured data like lab results and demographics but do not incorporate clinical notes, which are a rich source of individualised patient information.
WashU’s AI model leverages large language models (LLMs) to analyse unstructured data, outperforming conventional machine learning methods in predicting complications.
The model correctly identified 39 more at-risk patients per 100 complications than previous approaches.
Charles Alba, a graduate student and co-author of the study, emphasised the model’s adaptability. He said: ‘Foundation models can be diversified, making them generally more useful than specialised models. We fine-tuned our model for multiple tasks simultaneously and found it predicts complications more accurately than models trained specifically to detect individual complications.’
The model enhances prediction accuracy across various surgical outcomes by identifying shared risk factors across multiple complications.
Joanna Abraham, an associate professor of anaesthesiology at WashU Medicine, added: ‘This versatile model has the potential to be deployed across various clinical settings to predict a wide range of complications. Identifying risks early could become an invaluable tool for clinicians, enabling them to take proactive measures and tailor interventions to improve patient outcomes.’
Despite its promise, AI integration into clinical workflows presents challenges.
Additionally, trust in AI models from treating clinicians is crucial, requiring transparency and interpretability.
While current models function as ‘black boxes’, researchers are working on integrating interpretability mechanisms to enhance trust and usability.
Also, the study is restricted to data from Barnes-Jewish Hospital. Future research will include data from multiple healthcare systems to address variations in surgical documentation, terminology, and abbreviations, ensuring broader applicability.
- Meanwhile, a surgical conference in Ghent marked a significant milestone in AI-driven surgery by performing the first-ever operation with real-time AI commentary, according to VRT NWS.
Developed by the Orsi Academy (a Belgian centre for robotic and minimally invasive surgical training) in collaboration with Nvidia, the AI system analysed live footage and provided tailored explanations based on the expertise of the audience, including engineers and top surgeons.
The system allows real-time education without interrupting the surgical team, making it a promising tool for medical training and postoperative patient engagement.
The conference also showcased Belgium’s first remote robotic surgery. A surgical robot based in Shanghai was successfully controlled from Ghent, where it operated on a chicken leg with minimal delay.
