Do Machine Learning Approaches Perform Better Than Regression Models in Mapping Studies? A Systematic Review.

Area of Science:

• Health Economics and Outcomes Research
• Biostatistics
• Medical Informatics

Background:

• Mapping algorithms are crucial for synthesizing evidence from diverse health economic studies.
• Traditional regression models (RMs) have been widely used for developing these algorithms.
• Machine learning (ML) offers advanced computational approaches for algorithm development.

Purpose of the Study:

• To systematically review the implementation of machine learning (ML) in mapping studies.
• To quantify the performance improvements of ML approaches compared to traditional regression models (RMs).

Main Methods:

• A systematic literature search was performed across 12 databases up to December 2023.
• Studies applying ML for mapping algorithms were identified and data extracted.
• Goodness-of-fit indicators (e.g., R-squared, MAE) were compared between ML and RMs.

Main Results:

• Thirteen studies utilized both ML and RMs; Bayesian networks were the most frequent ML approach.
• ML demonstrated average improvements in goodness-of-fit: 0.007 (MAE), 0.004 (MSE), 0.058 (R-squared).
• Minor performance gains were observed with ML, particularly in mean-based comparisons.

Conclusions:

• The use of ML in developing mapping algorithms is increasing.
• ML generally offers minor improvements in goodness-of-fit over RMs.
• Challenges in ML interpretation, application, and external validation require further investigation.