📄 medicine

Predicting Positive Blood Cultures Early in Hospitalization Using Time Locked Electronic Health Record Data

This study demonstrates that a time-locked machine learning model using electronic health record data can effectively predict the risk of positive blood cultures within 24 hours of hospital admission, enabling the prioritization of patients for early intervention before final results are available.

Original authors: Radu Grovu, Hala Soomro, Nathan Gelinas, Michael Rossi, Sean F Monaghan

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Radu Grovu, Hala Soomro, Nathan Gelinas, Michael Rossi, Sean F Monaghan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Hospitals are places where the body's defenses are often already compromised, and when a patient develops a serious infection that spreads through the bloodstream, the stakes are incredibly high. To find the specific germ causing the illness, doctors draw blood and send it to a laboratory to grow the bacteria or fungi, a process known as a blood culture. This test is the gold standard for diagnosing life-threatening infections, yet it has a frustrating flaw: the vast majority of these tests come back negative, showing no growth at all. Sometimes the blood is truly clean, but often the test fails to catch the germ, or the sample is accidentally contaminated by skin bacteria during the draw. When a culture turns out positive, it is a critical moment that demands immediate attention, but because most results are negative, medical teams must treat every single sample with the same level of caution, which can strain resources and delay care for the few patients who truly need it. The challenge, then, is not to decide who should get a test, but to figure out, right after the blood is drawn, which samples are most likely to reveal a dangerous pathogen so that the right people can be prioritized for faster results and treatment.

A team of researchers at Brown University Health set out to solve this specific problem by building a computer model that acts like a smart filter for blood culture results. They looked at data from over 51,000 hospital admissions where blood cultures were taken within the first day of a patient's stay. The researchers trained a machine learning system to look at the information available in the patient's electronic health record at the exact moment the blood was drawn. This information included the patient's age, their medical history, whether they were in the intensive care unit, what medications they had recently received, and their current vital signs and blood test results. Crucially, the model was designed to be "time-locked," meaning it could only use facts that existed before the needle touched the vein, ensuring it would not use lab results before they were ready. The goal was to see if the computer could spot patterns in the data that human doctors might miss, predicting which patients were at the highest risk of having a true infection.

The study found that the model worked well at separating the likely infections from the negative results. In the group of patients the model had never seen before, it successfully identified a small group of high-risk individuals. While only about four percent of all the blood cultures in the study turned out to be positive, the model's predictions were far more accurate when focused on its top picks. Among the one percent of patients the model flagged as having the highest risk, nearly thirty percent actually had a positive culture. Even among the top five percent of predicted risks, the rate of positive results jumped to nearly twenty percent. This means the tool could effectively triple or quadruple the chance of finding a true infection if medical teams focused their immediate attention on these specific patients. The most powerful clue the model used was a simple piece of history: whether the patient had a positive blood culture in the past. Patients with a history of bloodstream infections were much more likely to have one again. However, even when the researchers removed this history from the model, the system still performed well, relying on other signals like recent hospital stays, the type of admission, and specific abnormalities in blood chemistry such as low platelet counts or high levels of waste products in the blood.

The researchers also tested how robust their findings were by changing the rules of the game. They checked if the model was just counting how many times a doctor ordered a test, or if it was truly learning about the patient's condition. They found that the model's success did not depend on how many tests were ordered, but on the actual clinical picture. They also confirmed that the model remained accurate even when they looked only at a patient's first hospital visit or when they excluded patients who died very quickly, suggesting the tool is stable and reliable across different scenarios. The study did not claim that this model should replace doctors or that it could predict infections in patients who never get a blood test drawn. Instead, the authors present it as a practical tool for "diagnostic stewardship," a concept that means using resources wisely. By using this computer score to prioritize which blood samples get looked at first by the lab or reviewed immediately by infectious disease specialists, hospitals could potentially speed up treatment for the sickest patients without wasting time on the many samples that will turn out to be negative.

This work represents a shift in how hospitals might manage the flood of data they generate every day. Rather than treating every blood culture as an equal mystery to be solved at the same pace, the study suggests that a digital assistant could help triage the workload, highlighting the cases where the odds of a serious infection are highest. The model is not a crystal ball, and the researchers emphasize that it needs to be tested in other hospitals to ensure it works in different settings with different patient populations. But the core finding is clear: by carefully analyzing the information available the moment a blood sample is taken, it is possible to identify a small group of patients who are far more likely to have a dangerous infection than the average person, allowing the medical system to focus its most urgent efforts where they are needed most.

Technical Summary: Predicting Positive Blood Cultures Early in Hospitalization Using Time Locked Electronic Health Record Data

Problem Statement
Blood cultures are the standard for diagnosing bloodstream infections and guiding antimicrobial therapy; however, the majority of cultures do not yield a pathogen. Contaminated or negative results can lead to unnecessary antibiotic use, extended hospital stays, and inefficient resource allocation. While previous prediction models have focused on determining whether to order a blood culture (often in the emergency department to avoid low-yield testing), there is a distinct operational gap in identifying which patients, after a culture has already been drawn, are at the highest risk for a true positive result. The authors posit that a model capable of enriching for positive cultures in this "post-order" window could facilitate diagnostic stewardship, allowing for prioritized laboratory follow-up, earlier infectious disease input, and targeted stewardship reviews before final results are available.

Methodology
The study utilized a retrospective cohort design using the MIMIC-IV (version 3.1) de-identified electronic health record dataset from Beth Israel Deaconess Medical Center.

Cohort: The study included 51,424 adult admissions where blood cultures were obtained within the first 24 hours of hospitalization. The dataset was split at the subject level to ensure no patient appeared in multiple partitions: 32,987 admissions for training, 8,186 for validation, and 10,251 for testing.
Prediction Anchor: The model was "time-locked" to the index culture time ( $t_0$ ). Only data available strictly before $t_0$ (including historical data and dynamic variables up to 24 hours prior) was used to predict the outcome. Post- $t_0$ measurements were excluded to prevent data leakage.
Outcome: The primary outcome was a pathogen-positive blood culture. Contaminants (e.g., coagulase-negative staphylococci) were treated as negative, with a tightened rule requiring multiple cultures with the same organism to classify an event as positive.
Modeling Approach: The authors employed XGBoost (gradient boosted trees) with fixed hyperparameters. Missing numeric values were handled natively by the model, while categorical missingness was encoded as a distinct category.
Validation: Performance was evaluated using AUROC, AUPRC, and Brier score. Confidence intervals were derived via clustered bootstrap resampling. Calibration was assessed using intercept and slope metrics.
Sensitivity Analyses: The study tested model robustness by removing workflow features, prior positive culture history, early deaths, measurement intensity features, and restricting the cohort to first admissions only.

Key Results

Cohort Characteristics: The test set comprised 10,251 admissions with a culture-positive prevalence of 4.04%. The median time from admission to culture was 7.6 hours.
Primary Model Performance: The primary model achieved an AUROC of 0.754 (95% CI, 0.726–0.778), an AUPRC of 0.157 (95% CI, 0.129–0.193), and a Brier score of 0.037. Calibration was good (intercept -0.108, slope 1.012).
Risk Enrichment: The model demonstrated significant ability to stratify risk. While the baseline prevalence was ~4%, the observed positivity rate in the top 1% of predicted risk was 30.1%, in the top 5% was 19.9%, and in the top 20% was 10.8%. The top 20% risk group captured 53.4% of all positive cultures.
Key Predictors:
- Primary Model: The strongest predictor was "Prior positive blood culture history" (Mean absolute SHAP: 0.561). Other significant predictors included time to culture, admission from the Emergency Department (ED), prior admissions count, BUN, prior procedures, medication orders, and various laboratory values (lymphocytes, bicarbonate, WBC, platelets, sodium).
- Refit Model (without prior history): When prior positive culture history was removed, the model retained predictive power (AUROC 0.662), relying on recent healthcare exposure (prior procedures, days since discharge), treatment context (antibiotic exposure, medication burden), and laboratory markers of acute illness (lower bicarbonate/platelets/sodium, higher WBC/lactate/INR).
Robustness: Performance remained stable when measurement intensity features were removed, suggesting results were not driven merely by the volume of testing. The model was also stable after excluding early deaths and restricting to first admissions.

Significance and Claims
The authors claim that this time-locked model successfully identifies a subgroup of adults, already cultured early in hospitalization, who have a substantially higher likelihood of pathogen-positive blood cultures compared to the general cohort.

Diagnostic Stewardship: The paper frames the model's utility not as a decision tool for ordering cultures, but as a prioritization tool for post-culture management. By flagging high-risk episodes, the model could enable earlier laboratory prioritization, pharmacy review, and infectious disease consultation.
Resource Allocation: The enrichment capability supports the efficient use of limited rapid diagnostic resources (e.g., molecular assays, sequencing) by directing them toward patients with the highest predicted yield.
Limitations: The authors explicitly note that the model is restricted to adults cultured within 24 hours of admission and originates from a single-center dataset (MIMIC-IV). They state that external and prospective validation are essential before clinical implementation, as culturing practices, contamination rates, and patient mixes vary across institutions. The study concludes that while the model shows promise for risk-based prioritization, further validation is required to assess its impact on turnaround time, treatment decisions, and resource utilization.

Technical Summary: Predicting Positive Blood Cultures Early in Hospitalization Using Time Locked Electronic Health Record Data

More like this