Lung infections like pneumonia are among the world’s top killers — but diagnosing them is notoriously hard.
Now, researchers at UC San Francisco have found a way to identify these infections in critically ill patients by pairing a generative AI analysis of medical records with a biomarker found in samples of lung fluid, in this case the expression of a gene that reduces inflammation.
In an observational study of critically ill adults, the combination made a correct diagnosis 96% of the time and distinguished between infectious and non-infectious causes of respiratory failure more accurately than clinicians in the intensive care unit. The authors estimated that this model, had it been available when the patients were admitted, could have cut inappropriate antibiotic use by more than 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., an associate professor of Medicine and senior author of the study, published Dec. 16 in Nature Communications. “We’re confident that it could lead to faster diagnosis and curtail the unnecessary use of antibiotics.”
An important feature of the model is the biomarker, which Langelier’s team developed in 2023. They found that a gene that tempers inflammation, called FABP4, is less expressed in infected lung cells than in healthy ones, and therefore could be used to help diagnose infection.
Calling all physicians to test in the clinic
The current study looked at data from two sets of critically ill patients: 98 were recruited before the COVID-19 pandemic, and most had bacterial infections; 59 were recruited during the pandemic, and most had viral infections, including COVID-19.
Before combing the two methods, researchers first tested each diagnostic method alone — FABP4 biomarker or AI — and found that each was correct about 80% of the time. The researchers then compared the models’ results with the diagnoses made by the doctors who had admitted the patients to the hospital’s intensive care unit.
Doctors prescribed antibiotics to treat pneumonia for most of these patients, while the biomarker-plus-AI model was much more judicious in assigning a diagnosis of pneumonia.
To further test the model’s accuracy, the team compared how the AI analyzed the medical records to how three different physicians who specialize in internal medicine and infectious diseases analyzed them. The AI was done by GPT4 on a privacy-protecting platform developed at UCSF.
Both got about the same number of diagnoses correct, but the AI gave more weight to radiology reports about the chest X-rays, while the 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, one of the first authors of the paper. “It shows how AI can complement the work physicians do.”
The team published their AI prompts in the paper and encouraged physicians to try them out on their own HIPAA-compliant AI platforms.
“Using this is unbelievably simple, you don’t have to be a bioinformatician,” said Hoang Van Phan, Ph.D., who is himself a bioinformatician as well as a first author of the paper.
The team is validating the model as a clinical test. Next, they will turn to sepsis, the most common cause of hospital death and one that is also famously hard to pin down.
Authors: Additional authors on the paper 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 of UCSF.
Funding: This work was supported by the National Institutes of Health (R01AI185511, NHLBI R35HL140026) and the Chan Zuckerberg Biohub.
Disclosures: No authors report financial or personal conflicts of interest.
