The Robust Classification Model Based on Combinatorial Features.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Disease phenotype characterization can be improved by analyzing combinatorial features.
• Existing classification methods may not fully capture complex disease patterns.

Purpose of the Study:

• To propose a novel method (CFC-CM) for constructing feature combinations and a classification model.
• To enhance disease identification accuracy by mining key feature relationships.

Main Methods:

• Developed a novel classification model (CFC-CM) using a modified k-top-scoring pair (M-k-TSP) algorithm.
• Iteratively tested feature relationships and selected discriminative feature pairs to build the model.
• Applied CFC-CM to gene expression and metabolomics datasets.

Main Results:

• CFC-CM demonstrated superior performance on nine public gene expression datasets compared to SVM, random forests, LASSO, elastic net, and M-k-TSP.
• Achieved high accuracy rates of 88.73±2.06% and 79.11±2.70% on metabolomics datasets for distinguishing specific diseases.
• Outperformed five other methods in classifying hepatocellular carcinoma vs. hepatic cirrhosis and acute kidney injury (AKI) vs. non-AKI samples.

Conclusions:

• Combinatorial features, inferred with an appropriate number of features, offer better disease identification than pairwise combinations.
• CFC-CM shows significant potential for more precise identification of complex diseases.
• The findings highlight the importance of feature combinations in understanding disease phenotypes.