2187. Prediction of Patient Outcome During Febrile Neutropenia Despite Anti-infective Treatment Using Machine Learning Algorithms
Abstract Background Clinical management of prolonged febrile neutropenia despite broad-spectrum empirical antibacterial treatment is a clinical challenge, as standard empirical treatment has failed and a broad spectrum of differential diagnoses has to be considered. Growing prevalence of multi-resistant bacteria and fungi has made a balanced choice of effective anti-infective treatment more difficult. A reliable prediction of complications could indicate options for treatment optimization. Methods We implemented a supervised machine learning approach to predict death or admission to intensive care unit within 28 days in cancer patients with prolonged febrile neutropenia (neutrophils < 500/mm3 and body temperature ≥ 38°C longer than 3 days). We analyzed highly granular retrospective medical data of the Cologne Cohort of Neutropenic Patients (CoCoNut) between 2008 and 2014. Random forest and 10-fold cross-validation were used for classification. The neutropenic episodes from 2014 were used for evaluation of prediction. Results In total, 927 episodes of prolonged febrile neutropenia (median age 52 years, interquartile range 42–62; 562/927 [61%] male; 390/927 [42%] acute myeloid leukemia; 297/927 [32%] lymphoma) with 211/927 (23%) adverse outcomes were processed. We computed 226 features including patient characteristics, medication, clinical signs, as well as laboratory results describing changes of state and interactions of medical parameters. Feature selection revealed 65 features with an area under the receiver operating characteristic curve (AUC) of 0.75. In the validation data set the optimized model had a sensitivity/specificity of 36% and 99% (AUC: 0.68; misclassification error: 0.12) and positive/negative predictive values of 89% and 88%, respectively. The most important features were albumin, age, and procalcitonin. Conclusion Structured granular medical data and machine learning approaches are an innovative tool that can be used in a retrospective setting for prediction of adverse outcomes in patients with prolonged febrile neutropenia. This study is the first important step toward clinical decision support based on predictive models in high-risk cancer patients. Disclosures All authors: No reported disclosures.