Multivariate binary classification of imbalanced datasets-A case study based on high-dimensional multiplex autoimmune...

Area of Science:

• Medical diagnostics and predictive medicine
• Machine learning in healthcare
• Biostatistics and data analysis

Background:

• Accurate patient classification is crucial for diagnosis and treatment.
• Class imbalance, where subgroups are small (e.g., rare diseases), complicates statistical classification.
• High error rates in minority classes are a common issue in imbalanced datasets.

Purpose of the Study:

• To investigate class imbalance effects on Random Forests and Powered Partial Least Squares Discriminant Analysis (PPLS-DA).
• To evaluate classifiers combined with imbalance compensation methods (sampling, cost-sensitive learning).
• To assess classifier performance using a scoring system on an autoimmune assay dataset.

Main Methods:

• Case study using a high-dimensional multiplex autoimmune assay dataset (Rheumatoid Arthritis vs. Systemic Lupus Erythematodes).
• Simulated class imbalance by reducing the majority class size.
• Comparison of Random Forests and PPLS-DA with and without sampling and cost-sensitive learning.

Main Results:

• Cost-sensitive learning approaches showed potential benefits for Random Forests in handling class imbalance.
• Performance evaluation utilized a scoring system to analyze classification outcomes.
• The study highlights the impact of imbalance on classifier error rates.

Conclusions:

• Cost-sensitive learning may be a valuable strategy for improving Random Forest performance with imbalanced medical data.
• Further research with diverse datasets and simulations is recommended.
• This work emphasizes the importance of addressing class imbalance in medical data analysis.