E385: Predicting post-operative infection in splenectomy patients with a novel machine learning model

Introduction: Hematologic changes following splenectomy can make it difficult to use standard metrics to differentiate between infection and a normal post-operative response. Heterogeneity in baseline laboratory values due to neoadjuvant chemotherapy, indication for splenectomy, or receipt of hyperthermic intraperitoneal chemotherapy (HIPEC) can further complicate assessment. The aim of this study was to develop a machine learning model to predict the development of a post-operative infection following splenectomy.

Methods: Patients who underwent splenectomy from 7/2015 to 12/2022 at a single quaternary center were identified. Demographic, comorbidity, laboratory (white blood cell [WBC], platelets, lymphocyte, and neutrophil counts/percentages), and culture data (urine, respiratory, and blood) were collected. Patients were divided into infected and non-infected cohorts by indication for splenectomy, and the post-operative course was limited to 14 days. Machine learning (XGBoost) was used to model the development of a post-operative infection based on all variables collected. Patients < 18 years of age or those who underwent splenectomy for hematologic disorders were excluded.

Results: A total of 921 patients were included in the study, with 645 in the training group (70%) and 276 in the validation group (30%). The median age was 60 years (IQR 47-69) and 466 (51%) were female. The majority of patients underwent splenectomy for pancreatic disease (n=395, 43%), followed by 28% (n=254) for benign indications/trauma, 18% (n=170) for other cancers, and 11% (n=102) as part of CRS/ HIPEC for peritoneal metastases. Median WBC levels were significantly different between the infected and non-infected group in the trauma cohort for 13 of 14 post-operative days (p< 0.01), while there was minimal difference between the cohorts for patients with pancreatic disease, and no difference at any time point for the other cancers and HIPEC patients (Figure 1). XGboost exhibited excellent prediction accuracy for post-operative infection, with an average area under the curve of 0.83, negative predictive value of 96%, positive predictive value of 46%, and overall accuracy of 94%.

Conclusions: By utilizing deep machine learning, we developed a novel algorithm to predict the development of a post-operative infection in splenectomy patients, regardless of the indication for splenectomy. In the future, this tool can help prevent costly imaging and premature treatment of potential infections by confirming the absence of an active infection.