Artificial Neural Network for the Fast Screening of Samples from Suspected Urinary Tract Infections.

Area of Science:

• Clinical diagnostics
• Artificial intelligence in medicine
• Urine analysis

Background:

• Urine microbial analysis is prone to contamination, causing false positives and delays.
• Digitalization and machine learning (ML) offer potential solutions for clinical decision support.
• Contamination leads to misdiagnosis and inefficient healthcare resource allocation.

Purpose of the Study:

• To investigate the use of a simple artificial neural network (ANN) for pre-identifying negative and contaminated urine specimens.
• To develop and evaluate a machine learning model for urine sample analysis.
• To reduce false-positive diagnoses and unnecessary microbial cultures.

Main Methods:

• Developed a multilayer perceptron (MLP) model using 8181 urine samples (cytology, dipstick, culture).
• Randomly split data 2:1 for training and testing; excluded low-importance variables.
• Utilized microbial and urine color/urobilinogen data as primary inputs for the model.

Main Results:

• The MLP model achieved a negative predictive value (NPV) of 96.5% and a positive predictive value (PPV) of 87.2%.
• Contaminated specimens were often misclassified as negative.
• Identified 0.82% of cultures as unnecessary microbial cultures (UMC).

Conclusions:

• ANN models reliably screen negative urine samples, aiding in the detection of significant bacteriuria.
• The model is effective for ruling out negative results but less so for confirming positive ones.
• Further improvements in accuracy and reduction of false negatives may be achieved by incorporating morphological data.