Comparative study of machine-and deep-learning based classification algorithms for biomedical Raman spectroscopy...

Area of Science:

• Biomedical spectroscopy
• Machine learning in diagnostics
• Deep learning for microbial identification

Background:

• Advancements in biomedical Raman spectroscopy (RS) are driven by machine learning (ML) and deep learning (DL) algorithms.
• A systematic comparison of ML and DL for biomedical RS is lacking due to limited open-source spectral data.

Purpose of the Study:

• To compare the performance of a typical ML algorithm (PLS-DA) and a DL algorithm (1D-CNN) for pathogenic microbe identification using Raman spectra.
• To evaluate algorithm performance across varying dataset sizes (100%, 75%, 50%, 25% of 12,000 spectra).

Main Methods:

• Utilized 12,000 Raman spectra from six microbial species.
• Compared partial least square-discriminant analysis (PLS-DA) and one-dimensional convolutional neural network (1D-CNN).
• Employed an 80% training and 20% testing data split for analysis.

Main Results:

• 1D-CNN achieved higher accuracy (95.25%) and AUC (0.997) than PLS-DA (89.42% accuracy, 0.979 AUC) with 100% of the data.
• PLS-DA outperformed 1D-CNN when using only 75%, 50%, and 25% of the Raman spectra.
• Both methods demonstrated reliance on the number of spectra and showed comparable interpretability of key spectral features (DNA, proteins).

Conclusions:

• Both ML and DL algorithms are valuable for Raman spectra identification.
• Algorithm selection should be application-dependent, considering data availability and desired accuracy.
• Further exploration of ML and DL is recommended for optimizing microbial identification via Raman spectroscopy.