Researchers at the University of California, San Francisco (UCSF) have developed a new approach to diagnosing severe lung infections, such as pneumonia, in critically ill patients. The method combines generative artificial intelligence (AI) analysis of medical records with a biomarker from lung fluid samples. This biomarker is based on the expression of the gene FABP4, which plays a role in reducing inflammation.
In an observational study involving critically ill adults, this combined method achieved a correct diagnosis rate of 96%. It was also more accurate than intensive care unit clinicians at distinguishing between infectious and non-infectious causes of respiratory failure. The researchers estimate that if this model had been available when patients were admitted, inappropriate antibiotic use could have been reduced by over 80%.
“We’ve devised a method that gives results much faster than a culture, and it could be easy to implement in the clinic,” said Chaz Langelier, M.D., Ph.D., associate professor of Medicine and senior author of the study published December 16 in Nature Communications. “We’re confident that it could lead to faster diagnosis and curtail the unnecessary use of antibiotics.”
The biomarker component was developed by Langelier’s team in 2023. They discovered that FABP4 is expressed at lower levels in infected lung cells compared to healthy ones, making it useful for diagnosing infection.
The study included two groups: one group of 98 patients recruited before the COVID-19 pandemic—most with bacterial infections—and another group of 59 patients recruited during the pandemic—most with viral infections such as COVID-19.
Each diagnostic method—the FABP4 biomarker or AI alone—was about 80% accurate individually. When used together, accuracy improved significantly. The researchers compared their model’s diagnoses with those made by hospital intensive care doctors and found that physicians prescribed antibiotics for most cases suspected as pneumonia, while the new model applied greater caution before assigning such diagnoses.
To further test accuracy, three specialist physicians analyzed patient records and their results were compared with those produced by GPT-4-based AI running on a privacy-protecting platform developed at UCSF. Both approaches achieved similar correct diagnosis rates; however, AI placed more emphasis on radiology reports while physicians focused on clinical notes.
“It was almost showing a cultural difference, if you can say that about an AI,” said Natasha Spottiswoode, M.D., DPhil, assistant professor of Medicine and first author. “It shows how AI can complement the work physicians do.”
The research team has published their AI prompts for others to use on HIPAA-compliant platforms. “Using this is unbelievably simple, you don’t have to be a bioinformatician,” said Hoang Van Phan, Ph.D., also a first author.
Currently being validated as a clinical test at UCSF, researchers plan next to adapt their model for sepsis diagnosis—a leading cause of hospital deaths known for its diagnostic difficulty.
Other authors include Emily Lydon, M.D.; Carolyn Calfee, M.D., MAS; Victoria Chu, M.D., MPH; Adolfo Cuesta, M.D., Ph.D.; Alexander Kazberouk, M.D., MBA; Natalie Richmond, M.D.; and Padmini Deosthale MS—all from UCSF. Funding came from the National Institutes of Health and Chan Zuckerberg Biohub. The authors disclosed no financial or personal conflicts of interest.
