Subsumption reduces dataset dimensionality without decreasing performance of a machine learning classifier.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• High-dimensional datasets are common in electronic health records.
• Hierarchical organization of features presents opportunities for dimensionality reduction.
• Subsumption leverages hierarchical relationships to generalize concepts.

Purpose of the Study:

• To investigate the efficacy of subsumption for reducing dimensionality in a disease dataset.
• To assess the impact of dimensionality reduction via subsumption on classification accuracy.
• To determine if subsumption is a viable strategy for high-dimensional, hierarchically structured data.

Main Methods:

• A neurological disease dataset with 293 features was used.
• Subsumption was applied iteratively to create eight datasets of decreasing dimensionality (293 to 11 dimensions).
• A Multi-Layer Perceptron (MLP) classifier was evaluated on all eight datasets.

Main Results:

• Dimensionality was successfully reduced from 293 to 11 features.
• Classification accuracy, precision, recall, and validation remained largely intact until the lowest dimensionality.
• Performance decline was observed only at the most reduced dimensionality.

Conclusions:

• Subsumption is a promising technique for reducing dimensionality in datasets with hierarchical feature organization.
• This approach can maintain classification performance while simplifying complex datasets.
• Clinical relevance: Applicable to high-dimensional electronic health record data structured by hierarchical ontologies.