Serum mass spectral fingerprints with machine learning for early discrimination of infectious pneumonia in the emergency department.

backgroundCommunity–acquired pneumonia is a major cause of emergency department visits, hospitalization, and death. In the emergency department, decisions to diagnose pneumonia and initiate antibiotics are typically guided by clinical assessment (symptoms and physical examination), chest imaging, and laboratory tests including inflammatory markers. However, discrimination between infectious pneumonia and non–infectious conditions remains only moderately accurate, and imaging is not always immediately available. There is therefore a need for simple, rapid point–of–care testing (POCT) that can screen for infectious pneumonia early in the evaluation. Serum metabolomics using liquid chromatography–mass spectrometry (LC–MS) is a potential POCT approach, but most prior studies have focused on relatively small panels of identified metabolites and have made limited use of unidentified spectral information. We therefore applied machine learning to full–scan serum mass spectral fingerprints, including unidentified peaks, to distinguish infectious pneumonia from non–infectious cases in the emergency department setting.

methodsWe conducted a single–center proof–of–concept observational study using a serum biobank from adult patients in a secondary–care ED in Japan. To evaluate the diagnostic models, we selected from this biobank 20 clearly non–infectious cases without gray–zone presentations and 20 cases of clinically diagnosed infectious pneumonia based on prespecified stringent clinical criteria. We performed LC–MS–based profiling to quantify metabolites and to acquire mass spectral fingerprints. Two machine–learning models were evaluated with stratified 5–fold cross–validation, and diagnostic performance for distinguishing these predefined groups was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity.

resultsWe analyzed 20 cases of infectious pneumonia and 20 non–infectious cases. The metabolite model achieved an AUC of 0.885 (95% confidence interval 0.754–0.985) for discriminating infectious pneumonia from non–infectious cases, whereas the mass spectral fingerprint model showed perfect discrimination in internal cross–validation (AUC = 1.000).

conclusionsIn this proof–of–concept study, machine–learning models trained on serum mass spectral fingerprints demonstrated promising performance in discriminating infectious pneumonia from non–infectious conditions in this selected cohort. This approach may serve as a foundation for future POCT applications, although larger multicenter external validation studies are warranted.