Abstract
BACKGROUND: Lower respiratory tract infections (LRTIs) are a leading cause of mortality worldwide and can be difficult to diagnose in critically ill patients, as non-infectious causes of respiratory failure can present with similar clinical features. [Figure: see text] [Figure: see text] METHODS: We developed a LRTI diagnostic method combining the pulmonary transcriptomic biomarker FABP4 with electronic medical record (EMR) text assessment using the large language model Generative Pre-trained Transformer 4 (GPT-4). We evaluated this approach in a prospective cohort of critically ill adults with acute respiratory failure from whom tracheal aspirate FABP4 expression was measured by RNA sequencing. Patients with LRTI or non-infectious conditions were identified using retrospective, multi-physician clinical adjudication. We then confirmed our findings by applying this method to an independent validation cohort of 115 adults with acute respiratory failure (Figure 1). Additionally, we compared GPT-4 diagnostic performance to physicians given the identical EMR information. [Figure: see text] RESULTS: In the derivation cohort, a combined classifier incorporating FABP4 expression and GPT-4–assisted EMR analysis achieved an AUC of 0.93 (±0.08) and an accuracy of 84%, outperforming FABP4 expression alone (AUC 0.84 ± 0.11) and GPT-4–based analysis alone (AUC 0.83 ± 0.07; Figure 2). By comparison, the primary medical team’s admission diagnosis had an accuracy of 72%. In the validation cohort, the combined classifier yielded an AUC of 0.98 (±0.04) and an accuracy of 96%. In comparison to human physicians, GPT-4 over-indexed on chest X-ray reads and under-emphasized notes from the clinical team (Figure 3). CONCLUSION: Integrating a host transcriptional biomarker with EMR text analysis using a large language model may offer a promising new approach to improving the diagnosis of LRTIs in critically ill adults. As a next step, we plan to extend this approach to other critical illness infectious syndromes, such as sepsis. DISCLOSURES: All Authors: No reported disclosures