Purpose: A German multicentre study BLOOMY was the first to use machine learning approach to develop mortality prediction scores for bloodstream infection (BSI) patients, but the scores have not been assessed in other cohorts. Our aim was to assess how the BLOOMY 14-day and 6-month scores estimate mortality in our cohort of 497 cases with BSI.
Methods: Clinical data, laboratory data, and patient outcome were gathered retrospectively from patient records. The scores were calculated as presented in the BLOOMY study with the exception in the day of the evaluation.
Results: In our cohort, BLOOMY 14-day score estimated death by day 14 with an area under curve (AUC) of 0.87 (95% Confidence Interval 0.80-0.94). Using ≥ 6 points as a cutoff, sensitivity was 68.8%, specificity 88.1%, positive predictive value (PPV) 39.3%, and negative predictive value (NPV) 96.2%. These results were similar in the original BLOOMY cohort and outweighed both quick Sepsis-Related Organ Failure Assessment (AUC 0.76) and Pitt Bacteraemia Score (AUC 0.79) in our cohort. BLOOMY 6-month score to estimate 6-month mortality had an AUC of 0.79 (0.73-0.85). Using ≥ 6 points as a cutoff, sensitivity was 98.3%, specificity 10.7%, PPV 25.7%, and NPV 95.2%. AUCs of 6-month score to estimate 1-year and 5-year mortality were 0.80 (0.74-0.85) and 0.77 (0.73-0.82), respectively.
Conclusion: The BLOOMY 14-day and 6-month scores performed well in the estimations of mortality in our cohort and exceeded some established scores, but their adoption in clinical work remains to be seen.
Keywords: Bacteraemia; Machine learning; Score; Sepsis.
© 2024. The Author(s).