Assessment of the risk of bacteremia in patients with hematologic malignancies in the emergency department: A comparative study between logistic regression and machine learning algorithms.

OBJECTIVE: To stratify the risk of bacteremia at the time of emergency department (ED) admission in patients with hematologic malignancies. To this end, we compared the performance of unsupervised and supervised machine learning algorithms with the classical multivariable logistic regression model. METHODS: We conducted a multicenter, international, retrospective cohort study including consecutive ED visits of adult patients with hematologic malignancies for whom blood cultures were obtained. The primary outcome was the prediction of bacteremia. The Bacteremia Objective Score System (BOSS)-1 used multivariable logistic regression, BOSS-2 applied K-means clustering, and BOSS-3 employed a support vector machine algorithm. Discriminative performance was assessed using sensitivity and specificity. External validation of the results was performed. RESULTS: The derivation cohort included 679 ED visits. Blood cultures tested positive in 88 cases (13%). BOSS-1 identified low-risk patients (3.1% bacteremia) with high sensitivity (94%; 95% CI, 87-97) but limited specificity (30%; 95% CI, 26-34). BOSS-2 better identified high-risk patients (29% bacteremia; sensitivity 93% [95% CI, 84-97], specificity 30% [95% CI, 26-34]). BOSS-3 classified all patients into low-risk (66.8%) or high-risk (33.2%) categories, without an intermediate group, unlike the other models, showing the best balance between sensitivity (61%; 95% CI, 59-64) and specificity (71%; 95% CI, 70-73) and yielding the most reproducible results in the external validation cohort. CONCLUSIONS: The BOSS algorithms enabled effective bacteremia risk stratification in the ED, with supervised machine learning models demonstrating the greatest potential clinical utility.