Using machine learning to develop smart reflex testing protocols.

Area of Science:

• Clinical diagnostics
• Laboratory medicine
• Machine learning applications

Background:

• Reflex testing enables secondary diagnostic tests from existing specimens.
• Current reflex testing relies on simple "if-then" rules, limiting its application.
• Complex ordering decisions often exceed the capabilities of rule-based systems.

Purpose of the Study:

• To propose and evaluate a machine learning-based approach for "smart" reflex testing.
• To predict ferritin test ordering based on complete blood count (CBC) testing.
• To compare machine learning performance against traditional rule-based reflex testing protocols.

Main Methods:

• Developed a machine learning model using deidentified patient data.
• Trained the model to predict ferritin test orders following CBC tests.
• Assessed model performance against actual ordering and rule-based approaches.

Main Results:

• Machine learning models showed moderate predictive performance for ferritin test ordering (AUC=0.731).
• Evaluated rule-based protocols lacked clinical feasibility due to insufficient agreement with actual ordering.
• Model deployment could prevent significant ferritin test ordering errors.

Conclusions:

• Machine learning offers a novel foundation for advanced reflex testing.
• This approach can enhance clinical diagnosis and optimize laboratory test ordering.
• Smart reflex testing holds potential for improved patient care and laboratory efficiency.