A machine learning strategy to mitigate the inappropriateness of procalcitonin request in clinical practice

Abstract

Aim: The aim of this study was to develop machine learning (ML) models to mitigate the inappropriate request of Procalcitonin (PCT) in clinical wards. Material and methods: We built six different ML models based on both demographical data, i.e., sex and age, and laboratory parameters, i.e., cell blood count (CBC) parameters, inclusive of monocyte distribution width (MDW), and C -reactive protein (CRP). The dataset included 1667 PCT measurements of different patients. Based on a PCT cut-off of 0.50 ng/mL, we found 1090 negative (65.4%) and 577 positive (34.6%) results. We performed a 70:15:15 train:validation:test splitting based on the outcome. Results: Random Forest, Support Vector Machine and eXtreme Gradient Boosting showed optimal performances for predicting PCT positivity, with an area under the curve ranging from 0.88 to 0.89. Conclusions: The ML models developed could represent a useful tool to predict PCT positivity, avoiding unusefulness PCT requests. ML models are based on laboratory tests commonly ordered together with PCT but have the great advantage to be easy to measure and low-cost.