PAM clustering algorithm based on mutual information matrix for ATR-FTIR spectral feature selection and disease...

Area of Science:

• Biomedical Spectroscopy
• Data Science in Healthcare
• Neurological Disorder Diagnostics

Background:

• Attenuated Total Reflectance Fourier Transform Infrared (ATR-FTIR) spectroscopy offers valuable data for diagnosing various pathologies, including neurological disorders.
• Selecting informative spectral biomarkers from large datasets is challenging due to data redundancy, hindering accurate disease discrimination.

Purpose of the Study:

• To develop a novel feature selection approach for identifying spectral biomarkers from ATR-FTIR data.
• To leverage redundant information within spectral data for improved disease prediction and interpretation.

Main Methods:

• A new feature selection method based on the Partition Around Medoids (PAM) algorithm was employed.
• A dissimilarity matrix derived from mutual information was used to group wavenumbers with similar dependence patterns.
• The medoid of each cluster, representing observed data points, was selected as a potential spectral biomarker.

Main Results:

• The proposed method successfully grouped wavenumbers based on their pairwise dependence patterns.
• The medoids derived from these clusters were found to be representative of the spectral data.
• The approach demonstrated effectiveness in discriminating between multiple sclerosis patients and healthy subjects using real-world data.

Conclusions:

• The Partition Around Medoids algorithm, utilizing mutual information, provides an interpretable method for spectral biomarker discovery.
• This approach facilitates the selection of robust spectral features for disease prediction in neurological disorders.
• The study highlights the potential of ATR-FTIR spectral data and advanced feature selection for clinical diagnostics.