Imbalanced target prediction with pattern discovery on clinical data repositories

Area of Science:

• Clinical Informatics
• Biostatistics
• Machine Learning in Healthcare

Background:

• Clinical data repositories (CDRs) hold potential for outcome prediction and risk modeling.
• Current methods often require complex study design and data handling, limiting direct use by clinical domain experts.
• There's a need for accessible tools to perform first-hand prediction and identify patterns in existing CDR data.

Purpose of the Study:

• To develop an interpretable pattern model for clinical data repositories.
• To enable clinical domain users to perform direct prediction without complex data handling.
• To uncover insightful patterns from imbalanced target variables for potential formal studies.

Main Methods:

• Proposed an interpretable, noise-tolerant pattern discovery model.
• Employed the geometric mean of sensitivity and specificity (G-mean) optimization for imbalanced targets.
• Developed a heuristic algorithm to address challenges in clinical research data.

Main Results:

• Pattern discovery showed competitive prediction performance on retrospective clinical datasets with imbalanced death rates (14.9% and 9.1%).
• Achieved statistically significant favorable G-means and F1-scores compared to logistic regression, Naïve Bayes, and decision trees.
• Consistently comparable prediction performance to the best achievable results without complex data processing.

Conclusions:

• Pattern discovery is robust and valuable for target prediction in clinical data repositories, handling imbalance and noise effectively.
• Provides agile and inexpensive insights through interpretable patterns for future formal studies.
• Facilitates direct application of predictive modeling in clinical practice.