Establishing reflex test rules for platelet fluorescent counting method using machine learning models on Sysmex...

Area of Science:

• Hematology
• Clinical Chemistry
• Machine Learning in Diagnostics

Background:

• The platelet fluorescent counting (PLT-F) method is a reflex test following the platelet impedance counting (PLT-I) method on Sysmex XN-series analyzers.
• Establishing optimized reflex test rules is crucial for efficient laboratory diagnostics.

Purpose of the Study:

• To develop and validate reflex test rules for the PLT-F method using multiple parameters from the Sysmex XN-series analyzer.
• To identify the most effective machine learning model for establishing these rules.

Main Methods:

• Evaluated baseline bias between PLT-F and PLT-I methods using 120 samples.
• Developed and tested reflex rules with 1256 samples using seven machine learning models.
• Utilized a 7:3 training-to-test set ratio and assessed model performance with various metrics.

Main Results:

• The PLT-F method showed strong correlation with the PLT-I method (r=0.998).
• The random forest model achieved the highest performance: 0.893 accuracy, 0.954 AUC, 0.771 F1 score, 0.719 recall, 0.831 precision, and 0.950 specificity.
• Mean cell volume was the most significant variable (15.09%) in the random forest model.

Conclusions:

• The random forest model is highly effective for establishing PLT reflex test rules.
• These rules, based on the PLT-F method and machine learning, enhance diagnostic efficiency on the Sysmex XN-series analyzer.