Artificial Intelligence-Based Prediction of Bloodstream Infections Using Standard Hematological and Biochemical Markers
Objective: Bloodstream infections (BSIs) require rapid identification to initiate timely antimicrobial therapy, yet blood culture-the current diagnostic gold standard-suffers from delayed results and limited sensitivity. This study aimed to develop an interpretable machine learning (ML) model using...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Galenos Yayinevi
2025-08-01
|
| Series: | Forbes Tıp Dergisi |
| Subjects: | |
| Online Access: | https://forbestip.org/articles/artificial-intelligence-based-prediction-of-bloodstream-infections-using-standard-hematological-and-biochemical-markers/doi/forbes.galenos.2025.57855 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Objective: Bloodstream infections (BSIs) require rapid identification to initiate timely antimicrobial therapy, yet blood culture-the current diagnostic gold standard-suffers from delayed results and limited sensitivity. This study aimed to develop an interpretable machine learning (ML) model using routine laboratory parameters to predict blood culture positivity.
Methods: A total of 1,972 adult patients who underwent complete blood count, C-reactive protein, procalcitonin (PCT), and blood culture testing at a tertiary hospital were retrospectively included. Three models-random forest, H2O automated ML, and an ensemble model-were developed and evaluated using standard classification metrics [area under the curve (AUC)-receiver operating characteristic (ROC), sensitivity, specificity, F1 score]. SHapley Additive exPlanations (SHAP) analysis was employed to enhance interpretability.
Results: The ensemble model yielded the best performance, achieving an AUC-ROC of 0.95, sensitivity of 0.78, specificity of 0.97, and F1 score of 0.84. External validation on an independent cohort confirmed the model’s generalizability (AUC-ROC: 0.85). SHAP analysis revealed that age and PCT were the most influential features with both statistical and clinical relevance. Basophil count, while ranked highest by SHAP, showed low sensitivity, highlighting the difference between algorithmic weight and bedside utility.
Conclusion: These findings support the integration of routine, readily available laboratory data into an explainable AI framework to accurately predict culture positivity. The model’s strong performance and interpretability suggest its potential application in clinical decision support systems to improve diagnostic stewardship, reduce unnecessary cultures, and optimize resource use in suspected BSI cases. |
|---|---|
| ISSN: | 2757-5241 |