Work

Predicting pathogen-positive blood culture episodes at two hours using electronic health record data

Year: 2026

Type: article

Abstract: Most blood cultures obtained early in hospitalization do not yield a pathogen. We evaluated prediction 2 h after the first blood-culture record using MIMIC-IV. Adults with an initial routine culture record within 24 h of hospital presentation were eligible if alive and still receiving care at prediction. The cohort included 145,914 encounters, 99.3% eventually admitted, and 6,404 pathogen-positive episodes (4.39%). Comparisons at 2, 4 and 6 h supported the earliest tested prediction time because later gains were small and uncertain. The primary XGBoost model was evaluated in 11,523 test encounters; 242 episodes were positive (2.10%). AUROC was 0.819 (95% CI 0.792–0.844), average precision 0.124 (0.096–0.167), and Brier score 0.0201. The top 5% of predicted risk contained 81 positives among 577 encounters, giving 14.0% positivity and 6.68-fold enrichment. Laboratory values, particularly lymphocyte percentage, contributed most to prediction. Secondary analyses assessed the added value of previous culture results. Mean predicted risk was 3.78%, compared with 2.10% observed. This 2-h model was able to concentrate pathogen-positive episodes in higher-risk groups, but calibration and transportability require further validation before clinical use.

Source: Scientific Reports

DOI: 10.1038/s41598-026-75707-6

Authors: Radu Grovu, Hala Soomro, Nathan Gelinas, Michael Rossi, Sean F. Monaghan

Institutions: Brown University, Rhode Island Hospital, Providence Health Care

Language: English

Topic: Machine Learning in Healthcare

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

SDG: Good health and well-being

Open Access status: gold

Estimated APC paid: $2,690