Beyond Comparing Machine Learning and Logistic Regression in Clinical Prediction Modelling: Shifting from Model...

Area of Science:

• Clinical prediction modelling
• Machine learning in healthcare
• Statistical modelling

Background:

• Supervised machine learning (ML) is increasingly used for clinical prediction, especially for binary outcomes from tabular data.
• Debate exists on ML's advantage over traditional logistic regression in this domain.
• ML excels in unstructured data, but performance gains in structured clinical data are inconsistent.

Purpose of the Study:

• To synthesize comparative studies and simulation findings on ML versus logistic regression in clinical prediction.
• To argue against a universally superior modelling approach.
• To identify factors influencing model performance and suggest avenues for improvement.

Main Methods:

• Literature synthesis of recent comparative studies.
• Review of simulation findings.
• Viewpoint argument based on synthesized evidence.

Main Results:

• No single modelling approach (ML or logistic regression) is universally best for clinical prediction with tabular data.
• Model performance is highly dependent on dataset characteristics (e.g., linearity, sample size, predictor count, class proportion) and data quality (completeness, accuracy).

Conclusions:

• Efforts should focus on improving data quality rather than increasing model complexity.
• Enhanced data quality is more likely to improve the reliability and real-world utility of clinical prediction models.