An interpretable machine learning approach to evaluate 30-day mortality risk in patients with community-onset bacteremia

Abstract

BACKGROUND: Machine learning (ML) techniques are increasingly being used in health outcome research to develop predictive models. However, ML models are often referred to as "black box models" because they lack interpretability. Our goal was to develop an ML model to predict mortality risk in patients with community-onset bacteremia. METHODS: We conducted a retrospective cohort study on 715 patients with bacteremia at a medical center in 2019. Model-agnostic methods were employed to visually explain the relationships between the predictors and the 30-day mortality risk. The model's performance was evaluated using the area under the receiver operating characteristic curve, calibration plots with the Brier score, accuracy, recall, precision, and F1 score. RESULTS: The top ten important predictors that significantly influenced the 30-day mortality prediction were the Pitt bacteremia score, septic shock, Charlson comorbidity index, length of stay in the ICU, neutrophil segment (%), age, neutrophil band (%), glucose, lymphocytes (%), and hemoglobin. The top three overall interaction strengths were septic shock, Charlson comorbidity index and Pitt bacteremia score, all of which significantly interacted with other predictors. CONCLUSION: ML revealed risk factors for 30-day mortality, including the Pitt bacteremia score, septic shock, age, pneumonia, and comorbidity, which also had multiple synergistic effects on 30-day mortality.

Ähnliche Arbeiten