UCSF researchers combine GenAI and genetic biomarkers to make pneumonia diagnosis faster and reduce antibiotic overuse.
Diagnosing pneumonia remains a critical challenge in intensive care units worldwide. Symptoms overlap with non-infectious respiratory failure. Cultures are slow and often inconclusive. Because of that, researchers at UC San Francisco (UCSF) applied GenAI to address pneumonia diagnosis faster and reduce antibiotic overuse. Their approach pairs AI-driven medical record analysis with a genomic biomarker. The goal is faster, more precise diagnosis, while reducing unnecessary antibiotic use.
The GenAI model analyzes unstructured clinical data using a privacy-preserving GPT-4 system. Whereas this reviews physician notes, lab results, and radiology reports. This AI assessment is combined with expression levels of the FABP4 gene. FABP4 is a biomarker linked to reduced inflammation during lung infections. Low gene expression signals infectious pneumonia rather than non-infectious causes.
Each diagnostic method alone reached roughly 80% accuracy. When combined, accuracy rose to 96%. The model outperformed ICU physicians in distinguishing infection types. Retrospective testing showed antibiotic use could have dropped by over 80%. GenAI addressed a key clinical challenge: rapid interpretation of complex, fragmented patient data. The system weighted imaging data more heavily than clinicians, revealing complementary reasoning patterns.
The study tested the model across pre-pandemic and COVID-era patient cohorts. This validated performance across bacterial and viral infections. Researchers published their AI prompts to encourage replication. The approach requires no advanced bioinformatics expertise. UCSF is now validating the system as a clinical test. Future GenAI applications include sepsis diagnosis, another data-intensive medical challenge.
This case matters because it illustrates how GenAI can enhance critical healthcare decision-making by rapidly synthesizing complex, fragmented clinical data. Beyond UCSF, it highlights a broader enterprise challenge: improving accuracy and efficiency in high-stakes, data-intensive environments where traditional methods are slow and error-prone. The case represents the problem of integrating diverse data sources—clinical notes, lab results, imaging, and biomarkers—into actionable insights. By combining AI-driven analysis with domain-specific signals, organizations can enhance diagnostic precision, reduce unnecessary interventions, and scale expert-level reasoning without requiring specialized expertise.
