Toward Informative Representations of Blood-Based Infrared Spectra via Unsupervised Deep Learning

Area of Science:

• Biomedical Engineering
• Computational Biology
• Spectroscopy

Background:

• Infrared molecular fingerprints of human blood contain complex information.
• Extracting meaningful diagnostic biomarkers from spectral data is challenging.
• Current methods may struggle with noise reduction and data dimensionality.

Purpose of the Study:

• To develop a deep learning model for low-dimensional representation of infrared blood spectra.
• To investigate the utility of this representation for lung cancer detection.
• To improve the accuracy and interpretability of spectral data analysis.

Main Methods:

• Utilized a fully convolutional denoising autoencoder for Fourier transform infrared (FTIR) spectroscopy data.
• Employed a bottleneck architecture and a custom loss function for noise reduction and information preservation.
• Applied the method to a case-control study for lung cancer detection.

Main Results:

• Successfully generated a low-dimensional latent space from complex FTIR spectra.
• Demonstrated effective noise reduction while retaining crucial molecular information.
• Achieved a 2.6 percentage point improvement in lung cancer detection accuracy.

Conclusions:

• Unsupervised deep learning offers a powerful approach for analyzing infrared molecular fingerprints.
• The developed autoencoder effectively compacts spectral data and identifies disease-associated variables.
• This methodology shows promise for enhancing diagnostic capabilities in medical research.