Probabilistic Prediction of Laboratory Test Information Yield

Area of Science:

• Clinical diagnostics
• Health informatics
• Medical laboratory science

Background:

• Repetitive laboratory diagnostic tests often yield minimal clinical information, contributing to increased healthcare costs and patient burden.
• Current methods for determining the necessity of repeated tests are often subjective and lack data-driven precision.

Purpose of the Study:

• To evaluate the predictability of stability in repeated laboratory diagnostic measurements using electronic health record (EHR) data.
• To develop a method for identifying low-yield repetitive tests to optimize diagnostic strategies and reduce healthcare expenditure.

Main Methods:

• Utilized probabilistic regression models to predict a distribution of plausible laboratory values based on pre-diagnostic EHR data.
• Developed 'stability' scores from predicted value distributions, allowing for customized definitions of stability based on clinical context.
• Assessed model performance in predicting test stability for various common laboratory diagnostics.

Main Results:

• High predictive accuracy for test stability was achieved for several key diagnostics, including 100% for platelets and 99% for albumin at 90% precision.
• The models demonstrated varying but significant sensitivity in predicting stability for other tests like hemoglobin (60%) and potassium (54%).
• These findings suggest a substantial fraction of repetitive tests could be safely reduced without compromising patient care quality.

Conclusions:

• Leveraging EHR data with probabilistic regression offers a feasible method for identifying and reducing low-yield repetitive laboratory tests.
• This data-driven approach enables personalized guidance for test utilization, enhancing efficiency and maintaining high standards of care.
• The study highlights the potential for significant cost savings and improved patient experience through optimized diagnostic testing strategies.