Predicting early readmission or postdischarge mortality in survivors of postoperative sepsis

Survivors of sepsis experience high rates of hospital readmission and postdischarge mortality. This study predicted 14-day unplanned readmissions or 14-day postdischarge mortality (postdischarge adverse outcomes) for survivors of postoperative sepsis or septic shock while identifying influential factors.

The National Surgical Quality Improvement Program (2018-2021) was queried for postoperative sepsis or septic shock during index admission. Univariable analysis compared patients with and without postdischarge adverse outcomes. Using variables from the index admission, Light Gradient-Boosting Machine and eXtreme Gradient Boosting models with and without synthetic minority oversampling (Synthetic Minority Oversampling Technique for Nominal and Continuous) predicted postdischarge adverse outcomes. SHapley Additive exPlanations were used for interpretation.

The cohort had 30,971 patients, with 13.4% having a postdischarge adverse outcome. Patients with postdischarge adverse outcomes had greater rates of comorbidities and complications during index admission (P < .05). eXtreme Gradient Boosting without Synthetic Minority Oversampling Technique for Nominal and Continuous had the highest F1 maximizing score (0.29) and area under the curve (0.65) and was further evaluated. The decile with the highest predicted risk represented 20% of all postdischarge adverse outcomes. SHapley Additive exPlanations identified National Surgical Quality Improvement Program mortality and morbidity probabilities, days from operation to discharge, sepsis present at the time of surgery, preoperative white blood cell counts, total work relative value units, American Society of Anesthesiologists class, operative time, and immunosuppression as the most influential factors across all predictions.

We identified unique perioperative predictors for postdischarge adverse outcomes in survivors of postoperative sepsis and septic shock. Despite techniques to address class imbalance, predictive performance remained modest. Future models should incorporate additional predictors that provide a more comprehensive representation of patient condition.