Countless Americans undergo surgical procedures annually. Post-surgery, mitigating complications such as pneumonia, thrombosis, and infections can make the difference between a successful recovery and an extended, distressing hospital stay — or even worse outcomes. More than 10% of surgical patients encounter such difficulties, which can result in prolonged stays in intensive care units (ICU), elevated mortality rates, and increased healthcare expenses. Timely recognition of patients at risk is essential, but accurately forecasting these risks continues to pose a challenge.
Recent developments in artificial intelligence (AI), especially large language models (LLMs), provide a promising resolution. A recent investigation spearheaded by Chenyang Lu, the Fullgraf Professor in computer science and engineering at the McKelvey School of Engineering and director of the AI for Health Institute at Washington University in St. Louis, examines the capability of LLMs to predict postoperative difficulties by scrutinizing preoperative evaluations and clinical notes. The findings, published online on Feb. 11 in npj Digital Medicine, demonstrate that specialized LLMs can significantly exceed traditional machine-learning techniques in anticipating postoperative hazards.
“Surgical procedures entail considerable risks and expenses, but clinical documentation is rich with valuable insights from the surgical team,” remarked Lu. “Our large language model, customized specifically for surgical documentation, facilitates early and precise predictions of postoperative complications. By proactively identifying risks, medical professionals can act sooner, enhancing patient safety and overall outcomes.”
Conventional risk assessment models have largely depended on structured data, which includes laboratory test results, patient demographics, and surgical specifics such as duration of procedure or the experience level of the surgeon. While this data is undoubtedly significant, it frequently lacks the depth present in a patient’s distinctive clinical narrative, which is encapsulated in the elaborate text of clinical notes. These documents encompass personalized descriptions of the patient’s medical background, present condition, and other elements that affect the probability of complications.
Lu, along with co-first authors Charles Alba and Bing Xue, both graduate students who collaborated with Lu during the research period, utilized specialized LLMs trained on publicly accessible medical literature and electronic health records. They subsequently refined the pretrained model on surgical notes to enhance predictions regarding surgical results. The resultant method — the first of its kind to analyze surgical notes and utilize them for predicting postoperative outcomes — can go beyond structured data to detect patterns in the patient’s condition that might otherwise be neglected.
Drawing on nearly 85,000 surgical notes and correlated patient outcomes from Barnes-Jewish Hospital gathered between 2018 and 2021, the team indicated that their model significantly surpassed traditional methods in anticipating complications. For every 100 patients experiencing a postoperative complication, the team’s new model accurately identified 39 additional patients who would have been overlooked by models relying on conventional natural language processing methods.
Apart from this impressive number of patients who could potentially have surgical complications identified early and alleviated, the study also highlights the strength of foundational AI models, which are built for multitasking and can be applied to diverse issues.
“Foundational models can be diversified, making them generally more advantageous than specialized models. In scenarios where multiple complications may arise, the model must be adaptable enough to forecast various outcomes,” noted Alba, who is also a graduate student in WashU’s Division of Computational & Data Sciences. “We refined our model for numerous tasks simultaneously and found that it predicts complications with greater accuracy than models designed specifically to detect single complications. This is logical since complications are often interrelated, so a unified foundational model benefits from shared insights about distinct outcomes and doesn’t need to be painstakingly adjusted for each one.”
“This adaptable model has the potential to be utilized across different clinical environments to forecast a broad spectrum of complications,” stated Joanna Abraham, an associate professor of anesthesiology and a member of the Institute for Informatics at WashU Medicine. “By early identification of risks, it could serve as an invaluable resource for clinicians, enabling them to implement proactive strategies and customize interventions to enhance patient outcomes.”
Alba C, Xue B, Abraham J, Kannampallil T, Lu C. The foundational capabilities of large language models in predicting postoperative risks using clinical notes. npj Digital Medicine, published online Feb. 11, 2025. DOI: https://www.nature.com/articles/s41746-025-01489-2
This research is funded by the Agency for Healthcare Research and Quality within the U.S. Department of Health and Human Services (R01 HS029324-02).
The post Foundation AI model predicts postoperative risks from clinical notes appeared first on The Source.