Machine learning based clinical prediction model for 1-year mortality in Sepsis patients with atrial fibrillation

Abstract

Atrial fibrillation (AF) emerges as a pivotal risk determinant for unfavorable outcomes in septic patients. Despite its recognized role, the enduring impact of AF on sepsis prognosis remains ambiguous. This investigation seeks to elucidate the connection between AF and both short and long-term outcomes in sepsis patients. Additionally, it aims to formulate a prognostic model for 1-year mortality utilizing pertinent clinical variables. A retrospective analysis encompassed sepsis patients admitted to Beth Israel Deacon Medical Center's intensive care unit. The evaluation encompassed the prevalence of AF and its influence on hospitalization duration, stays in the Intensive Care Unit (ICU), and mortality rates at distinct intervals. Propensity score matching was implemented to mitigate confounding factors. Machine learning techniques, including the Least Absolute Selection and Shrinkage Operator (LASSO) regression and random forest, were deployed for model development. AF exhibited a correlation with heightened mortality rates at 7 days, 28 days, and 1 year. The resultant predictive model demonstrated superior efficacy compared to prevailing clinical critical illness scores in forecasting mortality risk. Crucial predictors in the model included variables such as RDW, weight, age, BUN, lactate, temperature, MCHC, MBP, ALP, and hemoglobin. AF emerges as a substantial peril for adverse outcomes in sepsis patients. The risk model, encompassing pertinent clinical variables, outperformed existing clinical critical illness scores in mortality prediction. This model furnishes valuable insights for risk stratification, augmenting prognostic precision in sepsis patients with concomitant AF.