A deep learning solution to recommend laboratory reduction strategies in ICU.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Decision Support

Background:

• Laboratory testing generates vast amounts of data, contributing to healthcare costs and potential patient burden.
• Optimizing laboratory test ordering is crucial for efficient and effective patient care.

Purpose of the Study:

• To develop a machine learning model predicting laboratory test results.
• To create a strategy for reducing unnecessary laboratory tests using spatial-temporal correlations.
• To assist physicians in identifying which laboratory tests can be omitted.

Main Methods:

• A global prediction model was developed, treating laboratory testing as sequential decisions with contextual information.
• The model was validated on the MIMIC III critical care database (38,773 patients, 4,570,709 observations).
• Deep learning techniques were employed for real-time recommendations and prediction of lab test properties (values, normality, transitions).

Main Results:

• The best model achieved a 20.26% reduction in laboratory tests.
• On a hold-out dataset, the model predicted normality/abnormality with 98.27% accuracy (AUC 0.9885) on reduced tests.
• The model outperformed baseline greedy models, demonstrating a robust reduction strategy.

Conclusions:

• Spatial and temporal correlations in laboratory tests can optimize reduction policies.
• The proposed method offers an iterative prediction approach for dynamic laboratory test reduction.
• This machine learning model assists physicians in judiciously omitting laboratory tests.