Clinical prediction models for febrile neutropenia and its outcomes: a systematic review

Febrile neutropenia (FN) is a life-threatening complication of chemotherapy. Although practice guidelines suggest the use of existing prediction models when making decisions to prevent and treat FN, recent evidence suggests that these models are limited in their discriminative ability. This study aims to systematically review and critically evaluate the recent literature to assess the question: what evidence-based clinical prediction models can be used to predict FN or its outcomes?

PubMed, EMBASE, Web of Science, and SCOPUS were searched for primary journal articles that developed or validated models that predicted FN or outcomes in patients with FN. Risk of bias was critically evaluated using the Prediction model Risk of Bias Assessment Tool (PROBAST).

Five thousand nine hundred nineteen articles were identified, of which 90 met inclusion criteria. Twenty-five studies predicted FN, and 65 studies predicted outcomes in patients with FN, including 28 that predicted mortality, 35 that predicted microbiological outcomes, and 35 that predicted other complications. Eight studies used machine learning methods in their development, and few studies were externally validated. All 90 studies were graded as high risk of bias using PROBAST.

Prediction models for FN and its outcomes demonstrate promising discriminatory ability; however, several limitations have prevented these from translating clinically. These limitations include variable FN definitions, high ROB in current models, limited external validation, and heterogeneous cohorts. Future work is needed to further develop and validate robust, well-evidenced models that can translate into clinical practice. This may best be achieved through machine learning and electronic medical record integration.