Making complex prediction rules applicable for readers: Current practice in random forest literature and...

Area of Science:

• Medical and Health Sciences
• Computational Biology
• Biostatistics

Background:

• Prediction rules are essential for clinical decision-making.
• Simple prediction rules (e.g., logistic regression) are generally applicable.
• Complex prediction rules from machine learning pose challenges for reader application.

Purpose of the Study:

• To survey the applicability of prediction rules developed using the random forest algorithm.
• To identify barriers to the application of complex prediction rules.
• To provide recommendations for improving the applicability of such rules.

Main Methods:

• Survey of articles published in PLOS ONE (2014-2015) in medical and health sciences.
• Focused on prediction rules developed using the random forest algorithm.
• Assessed rule applicability and reproducibility.

Main Results:

• Only 2 out of 30 surveyed papers presented directly applicable prediction rules.
• Information for 8 additional rules was obtained by contacting authors.
• Author non-response and other issues hampered rule applicability in most cases.

Conclusions:

• Complex prediction rules, particularly those from machine learning, often lack direct applicability.
• Reproducibility and clear reporting are crucial for ensuring prediction rule usability.
• Recommendations are provided to enhance the applicability of future prediction rules.