Diagnostic performance of machine learning models using cell population data for the detection of sepsis: a comparative study

To compare the artificial intelligence algorithms as powerful machine learning methods for evaluating patients with suspected sepsis using data from routinely available blood tests performed on arrival at the hospital. Results were compared with those obtained from the classical logistic regression method.The study group consisted of consecutive patients with fever and suspected infection admitted to the Emergency Department. The complete blood counts (CBC) were acquired using the Mindray BC-6800 Plus analyser (Mindray Diagnostics, Shenzhen, China). Cell Population Data (CPD) were also recorded. The ML and artificial intelligence (AI) models were developed; their performance was evaluated using several indicators, such as the area under the receiver operating curve (AUC), calibration plots and decision curve analysis (DCA).Overall, all the tested approaches obtained an AUC>0.90. The logistic regression (LR) performed well compared to the ML/AI models. The naïve Bayes and the K-nearest neighbour (KNN) methods did not show good calibration properties. The multi-layer perceptron (MLP) model was the best in terms of discrimination, calibration and clinical usefulness.The best performance in the early detection of sepsis was achieved using the ML and AI models. However, external validation studies are needed to strengthen model derivation and procedure updating.