A diagnostic approach integrating a host immune biomarker with large language model (LLM) analysis of EHR significantly improves the diagnosis of lower respiratory tract infection in critically ill adults. 

By Ada Enesco


RT’s Three Key Takeaways:

  1. AI–biomarker integration: Combining the host immune biomarker FABP4 with GPT-4 analysis of electronic health records significantly improved diagnosis of lower respiratory tract infection in critically ill adults compared with either method alone.
  2. Superior diagnostic performance: The integrated model achieved high accuracy (84%–96%) and AUC values (0.93–0.98), outperforming standalone testing and standard clinician assessments in both primary and validation cohorts.
  3. Implications for antibiotic stewardship: By improving early identification of true infections when microbiology is inconclusive, this approach could help reduce unnecessary antibiotic use and support better ICU decision-making.


A novel diagnostic approach integrating a host immune biomarker with large language model (LLM) analysis of electronic health records significantly improves the diagnosis of lower respiratory tract infection in critically ill adults, according to new research. The findings highlight a potential strategy to reduce diagnostic uncertainty and inappropriate antibiotic use in intensive care settings. 

Diagnosing lower respiratory tract infection remains a major clinical challenge, particularly in the ICU, where infectious and non-infectious respiratory conditions often present similarly. Traditional microbiological testing frequently fails to identify a causative pathogen, leading clinicians to rely on empiric antibiotics, with well-recognized risks including Clostridioides difficile infection and antimicrobial resistance. 

Host transcriptional biomarkers offer a complementary approach by reflecting the body’s immune response rather than pathogen detection alone. In this study, researchers focused on FABP4, a single-gene pulmonary transcriptomic biomarker suitable for rapid molecular testing platforms already used in clinical practice. FABP4 expression was combined with LLM-based analysis of free-text electronic medical record (EMR) data using GPT-4. 

In a cohort of critically ill adults, the combined diagnostic classifier achieved an area under the receiver operating characteristic curve (AUC) of 0.93 and an overall accuracy of 84%. This performance exceeded that of FABP4 testing alone (AUC 0.84) and LLM-based EMR analysis alone (AUC 0.83). By comparison, the admitting clinical team’s diagnosis had an accuracy of 72%, highlighting the added value of the integrated approach. 

Importantly, the model’s performance was replicated in an independent validation cohort, where diagnostic accuracy rose to 96% with an AUC of 0.98. These findings suggest that combining structured molecular data with unstructured clinical narratives can meaningfully enhance diagnostic confidence in complex ICU cases. 

For clinicians, this approach may be particularly valuable when microbiological results are inconclusive or delayed. By improving early identification of true infection, and potentially distinguishing infectious from non-infectious causes of respiratory failure, the model could support more judicious antibiotic prescribing. 

While the study does not establish causality or directly assess patient outcomes, it provides proof of concept that integrating host biomarkers with AI can outperform either modality alone. Further prospective studies will be needed to evaluate real-world implementation, turnaround time, and impact on antimicrobial stewardship. 


Reference

Phan HV et al. Integrating a host biomarker with a large language model for diagnosis of lower respiratory tract infection. Nat Commun. 2025;16(1):10882.

This article was originally published by EMJ and was made available under the terms of the Creative Commons Attribution-Non Commercial 4.0 License.