Integrating a host transcriptomic biomarker with a large language model for diagnosis of lower respiratory tract infection

Abstract

AbstractLower respiratory tract infections (LRTIs) are a leading cause of mortality worldwide. Despite this, diagnosing LRTI remains challenging, particularly in the intensive care unit, where non-infectious respiratory conditions can present with similar features. Here, we tested a new method for LRTI diagnosis that combines the transcriptomic biomarkerFABP4with assessment of text from the electronic medical record (EMR) using the large language model Generative Pre-trained Transformer 4 (GPT-4). We evaluated this methodology in a prospective cohort of critically ill adults with acute respiratory failure, in which we measured pulmonaryFABP4expression and identified patients with LRTI or non-infectious conditions using retrospective adjudication. A diagnostic classifier combiningFABP4and GPT-4 achieved an area under the receiver operator curve (AUC) of 0.92 ± 0.06 by five-fold cross validation (CV), outperforming classifiers based onFABP4expression alone (AUC 0.83) or GPT-4 alone (AUC 0.84). At the Youden’s index within each CV fold, the combined classifier achieved a mean sensitivity of 92% ± 7%, specificity of 90% ± 17% and accuracy of 91% +/- 8%. Taken together, our findings suggest that combining a host transcriptional biomarker with interpretation of EMR data using artificial intelligence is a promising new approach to infectious disease diagnosis.