Rice University algorithm interprets spectra more accurately

01 May 2025

Improved machine learning identifies features that existing analysis can miss.

A project at Rice University has developed a new machine learning (ML) algorithm intended to improve the identification of biomarkers in optical spectra.

As reported in ACS Nano, the algorithm could potentially enable faster and more precise medical diagnoses and sample analysis.

"Machine learning methods have gained prominence in spectral analyses, efficiently unveiling analyte compositions," said the project in its published paper.

"However, these methods still face challenges in interpretability, particularly in generating clear feature importance maps that highlight the spectral features specific to each class of data. These limitations arise from feature noise, model complexity, and the lack of optimization for spectroscopy."

The Rice University solution is termed Peak-Sensitive Elastic-net Logistic Regression, or PSE-LR, a method tailored for spectral analysis. PSE-LR includes computational steps intended to more accurately detect spectral peaks without losing them in surrounding data, via three particular "regularization" operations.

Regularization stops ML routines from building too much on noise and outliers in its training data, encouraging the model to stay as simple as possible and nearer to baseline reality. In PSE-LR, this approach improves the algorithms ability to detect spectral differences arising from composition changes or other effects, parsing light-based data for subtle signals that are usually hard to pick up on using traditional methods.

Focusing on the important parts of the signal

"Most models either miss the tiny details or are too complex to understand," commented Ziyang Wang of Rice University's Sensing, Characterization, and Optoelectronics (SCOPE) Lab. "Our algorithm was designed to focus on the most important parts of the signal, the peaks that matter most."

PSE-LR can accurately classify different samples but is also transparent in its decision-making, something that many advanced ML models are not particularly good at, noted the project team. It will deliver a feature importance map that highlights exactly which parts of the spectrum contributed to a classification decision, which makes results easier to interpret, verify and act on.

In trials testing PSE-LR against other ML models, the new algorithm showed improved performance when detecting ultra-low concentrations of the SARS-CoV-2 spike protein in fluid samples, identifying neuroprotective solutions in mouse brain tissue, classifying Alzheimer’s disease samples and distinguishing between tungsten-based 2D semiconductor materials.

As well as improved spectral characterization of samples, PSE-LR could also spur the development of improved nanodevices and biosensors for new diagnostic methods, commented the project.

"These findings could help transform medical diagnostics and materials science, bringing us closer to a world where smart technologies help detect and respond to health problems faster and more effectively," said Ziyang Wang.