A fully interpretable machine learning model for increasing the effectiveness of urine screening.

Area of Science:

• Clinical Microbiology
• Medical Diagnostics
• Data Science in Healthcare

Background:

• Current urine culture screening methods are either too simple or overly complex.
• Limitations include reduced effectiveness of simple methods and lack of interpretability in "black box" machine learning models.
• There is a need for effective, interpretable screening to reduce workload, cost, and result turnaround time.

Purpose of the Study:

• To develop and validate an effective and interpretable screening method for identifying negative urine samples prior to urine culture.
• To overcome the limitations of existing screening approaches by utilizing transparent decision tree models.
• To reduce the workload and costs associated with microbiology laboratory processes.

Main Methods:

• Analysis of 15,312 samples from 10,534 patients using Sysmex UF-1000i automated analyzer data and clinical features.
• Development and application of decision tree (DT) models, with and without a lookahead strategy, for sample classification.
• Focus on creating transparent, logical rules for easy interpretation by medical professionals.

Main Results:

• The best performing DT model achieved 94.5% sensitivity in classifying negative samples.
• The model utilized age, bacteria, mucus, and two scattering parameters for classification.
• An additional 16% reduction in laboratory workload was achieved, with an estimated annual financial impact of €40,000.
• Identified logical rules demonstrated scientific rationale consistent with existing literature.

Conclusions:

• The study presents an effective and interpretable screening method for urine culture using automated analyzer data.
• The decision tree model offers transparency, fostering trust and facilitating real-world implementation in microbiology laboratories.
• This approach addresses the limitations of complex "black box" models and simple screening tools.