A random forest algorithm-based approach to capture latent decision variables and their cutoff values

Area of Science:

• Clinical Pathology
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Reference intervals (RIs) and clinical decision limits (CDLs) are crucial for interpreting laboratory results but are known to vary and possess flaws.
• Existing methods for establishing RIs and CDLs may not capture all clinically relevant thresholds, particularly those influencing patient outcomes.

Purpose of the Study:

• To develop and validate a novel methodology using a random forest algorithm to explore and identify latent clinical decision limits (latent CDLs) within established reference intervals.
• To apply this methodology to clinical laboratory data to identify latent CDLs associated with hospital length of stay (HLOS) in patients undergoing surgery.

Main Methods:

• A random forest algorithm was employed, utilizing phenotype transformation of independent variables to capture latent decision variables and their cutoff values.
• The methodology was applied to clinical laboratory data (blood, urine) from admitted patients, focusing on identifying latent CDLs for HLOS based on disease conditions.
• Specific attention was given to free Thyroxine (T4) levels as a potential indicator.

Main Results:

• Five distinct free Thyroxine (T4) cutoff values (1.16, 1.19, 1.2, 1.23, 1.25 ng/dL) were identified that predicted longer HLOS in patients with tachyarrhythmia.
• These identified cutoff values fall within the standard estimated RIs and hospital-determined RIs, suggesting limitations in current interval definitions.
• T4 levels above most of these cutoff values (except 1.16 ng/dL) were significantly associated with longer HLOS, indicating their clinical relevance despite being within normal RIs.

Conclusions:

• The proposed methodology effectively identifies latent CDLs that may indicate risks for prolonged hospital stays, even when values are within conventional RIs.
• These findings suggest that specific thresholds of free Thyroxine (T4) within RIs could serve as early alerts for longer HLOS in certain patient populations.
• Clinical experts should consider these latent CDLs for potential integration into electronic medical records to proactively manage patient risks upon admission.