Developing a machine learning prediction algorithm for early differentiation of urosepsis from urinary tract...

Area of Science:

• Biomedical engineering
• Clinical diagnostics
• Machine learning in healthcare

Background:

• Urosepsis poses significant risks, with early detection critical for reducing mortality and morbidity.
• Current diagnostic methods like blood cultures have limitations in sensitivity and speed.
• There is a need for rapid and accurate diagnostic tools for urosepsis.

Purpose of the Study:

• To develop a machine learning model for early urosepsis prediction using biomarkers.
• To identify key biomarkers for accurate and timely diagnosis of urosepsis.
• To improve clinical decision-making in managing urosepsis.

Main Methods:

• Retrospective analysis of 157 urosepsis and 417 urinary tract infection patients.
• Collection of laboratory data including procalcitonin, D-dimer, and C-reactive protein.
• Development and validation of machine learning models, including artificial neural networks (ANNs), using an 80/20 data split.

Main Results:

• Six machine learning models demonstrated over 80% accuracy in predicting urosepsis.
• An artificial neural network model incorporating the top eight biomarkers achieved the highest accuracy (92.9%) and AUC (0.946).
• The Gini importance ranking method effectively filtered significant variables for model optimization.

Conclusions:

• A data-driven predictive model using eight key biomarkers can accurately predict urosepsis.
• This model facilitates rapid and precise clinical diagnoses, aiding timely intervention.
• Biomarker-based machine learning offers a promising approach for early urosepsis detection.