Gram Stain Prediction with Machine Learning Techniques Using Biochemical Parameters in ICU Patients with Urinary Tract Infections

Purpose: The aim of this study was to develop a useful algorithm based on complete blood count (CBC), urinalysis, and biochemical parameters that could be an alternative to gram staining in the prediction of UTI and the determination of initial antibiotic treatment in ICU patients. Methods: All the specimens included in the study were obtained from ICU patients and were subjected to gram staining in the laboratory. Simultaneously, CBC, urinalysis, and biochemical tests were performed for each specimen. A classification based on biochemical parameters was performed for the estimation of gram-negative and gram-positive bacteria, as an alternative to gram staining. Results: Classification was achieved using multiple classification systems including Artificial Neural Networks (ANN), Support Vector Machine (SVM), the K-Nearest Neighbors (KNN), and Decision Tree Language (DTL) and the best classification performance was achieved by ANN, with an accuracy of 84.6%, sensitivity of 88.5%, and specificity of 73.5%. Conclusion: The high specificity and accuracy of the algorithm indicated that this algorithm can be effectively used in the selection of empirical antibiotic treatment for ICU patients with UTI and can provide more advanced and technological opportunities by combining laboratory parameters with machine learning techniques.