Interpretable Classification of Bacterial Raman Spectra With Knockoff Wavelets.

Area of Science:

• Biomedical Signal Processing
• Machine Learning in Healthcare
• Spectroscopy Analysis

Background:

• Machine learning models, including deep neural networks, are prevalent in biomedical signal analysis for pattern detection and prediction.
• Model interpretability remains a significant challenge, particularly in high-stakes medical applications like bacterial infection identification.

Purpose of the Study:

• To evaluate an interpretable logistic regression model against complex neural networks for bacterial infection identification using fast Raman spectroscopy data.
• To demonstrate that a simpler, transparent model can achieve comparable accuracy to more complex methods.

Main Methods:

• Utilized fast Raman spectroscopy data for bacterial infection analysis.
• Employed wavelet features with clear chemical interpretations.
• Implemented controlled variable selection using knockoffs to ensure predictor relevance and non-redundancy.

Main Results:

• The logistic regression model achieved accuracy comparable to neural networks.
• The selected wavelet features provided intuitive chemical interpretations.
• The variable selection method ensured predictors were relevant and non-redundant.

Conclusions:

• Interpretable logistic regression models with carefully selected features can match the performance of complex neural networks for biomedical signal analysis.
• The proposed approach, leveraging wavelet features and controlled variable selection, offers a transparent and broadly applicable alternative for interpretable machine learning in signal processing.
• This method is particularly valuable for applications requiring high-stakes decision-making, such as identifying bacterial infections.