Integrative nearest neighbor classifier for block-missing multi-modality data.

Area of Science:

• Biomedical data analysis
• Machine learning in healthcare
• Multi-modality data classification

Background:

• Biomedical classification often uses multi-modality data (e.g., omics, brain scans) for improved performance.
• Block-missing data, where entire modalities are absent in samples, is common due to high costs.
• Existing methods struggle with block-missing data, often requiring data deletion or imputation.

Purpose of the Study:

• To develop a novel classifier for block-missing multi-modality data.
• To effectively utilize all available information without data deletion or imputation.
• To improve classification accuracy in the presence of missing data modalities.

Main Methods:

• Introduced the integrative nearest neighbor (INN) classifier, a weighted nearest neighbors approach.
• INN adaptively weights training samples by minimizing worst-case estimation error.
• The method processes test data points using all available training information.

Main Results:

• INN outperforms standard weighted nearest neighbors classifiers using only complete samples or available modalities.
• INN shows superior performance compared to imputation methods, even with missing not at random data.
• Theoretical studies and a real-world Alzheimer's disease neuroimaging application validate INN's effectiveness.

Conclusions:

• The integrative nearest neighbor (INN) classifier is an effective solution for block-missing multi-modality data in biomedical classification.
• INN offers a robust alternative to data deletion and imputation, enhancing classification accuracy.
• This approach holds significant potential for applications with complex, incomplete biomedical datasets.