Enterprise AI Analysis

This retrospective study was conducted in the emergency departments of three hospitals in Taiwan. Data from adults with suspected bacterial infections were collected, including complete blood count, white blood cell differential count, and cell population data. A gradient boosting model (Catboost) was developed to classify nonbacteremia, Gram-negative and Gram-positive bacteremia. We evaluated the model through discrimination and calibration. The study included 28,503 samples in the CMUH development cohort, 15,801 in the CMUH validation cohort, 2632 in the WMH cohort, and 3811 in the ANH cohort. Each final laboratory report of CBC, DC, and CPD from a single specimen was considered one independent replicate. To minimize the risk of data leakage, samples from the same patient were assigned exclusively to either the training or the test set, regardless of whether they were collected during a single ED visit or across multiple visits. This approach mitigates bias from repeated measurements. Model performance was further validated across three independent hospital cohorts to demonstrate reproducibility.

The ML model achieved AUROC values ranging from 0.861 to 0.869 and AUPRC values from 0.325 to 0.415 for Gram-negative bacteremia. For Gram-positive bacteremia, the AUROC values ranged from 0.759 to 0.798 and the AUPRC values from 0.079 to 0.093. The development cohort achieved the highest sensitivity, at 0.814, whereas the WMH cohort had the highest specificity, at 0.745. The model's performance in predicting Gram-positive bacteremia was relatively suboptimal, with low AUROC and AUPRC values. Notably, the model's performance in predicting A. baumannii infection was markedly poor. Calibration plots for Gram-negative bacteremia prediction revealed generally good alignment, while for Gram-positive bacteremia prediction, the calibration plots revealed substantial deviation from the diagonal line, indicating poor calibration.

This study highlights the potential of CPD as one of the valuable inputs in bacteremia prediction, offering more detailed insights into patient immune responses. Our model, which integrates CBC, DC, and CPD, demonstrated robust potential for the early prediction of Gram-negative bacteremia in ED settings. Notably, results from both our three-class classifier and one-vs-rest binary classifiers consistently demonstrated that Gram-negative bacteremia can be effectively identified. In contrast, distinguishing Gram-positive bacteremia remains challenging, even when using dedicated binary models. This difficulty may be partly attributed to the overlapping distributions of laboratory features between Gram-positive and Gram-negative infections. Future studies should include more thorough investigations of the differences between Gram-positive and Gram-negative bacteria and further deployment of these models in clinical settings to assess their actual impact, such as patient outcomes, changes in patterns of antibiotic use, and whether they enhance patient safety.