Deep learning mitigates batch effects in Raman spectroscopy

Abstract

Abstract Raman spectroscopy is widely used for its ability to provide rich molecular-level information from biological samples. However, the high sensitivity of Raman spectra can also unintentionally encode extraneous information, such as patient identity, instrument-specific signatures, environmental conditions, and sample handling protocols. This hidden structure, referred to as the batch effect, poses increasing challenges in machine learning applications, where it can lead to model overfitting, poor generalization, and even privacy risks when datasets are shared. In this work, we investigate the presence and impact of the batch effect in a Raman spectroscopy dataset derived from glioma tumor samples. We show that deep learning can accurately predict patient identity from individual spectra, revealing strong bias patterns in the data. To mitigate this issue, we explore several approaches and introduce an algorithmic framework based on adversarial deep learning. Our method effectively reduces the influence of confounding sample-specific signals while preserving information relevant to tumor classification. These findings highlight the need for careful bias control in Raman-based AI pipelines and provide tools to support more robust applications in biomedical spectroscopy.

The authors' abstract, as published at the source. Scientific Reports, 2026 · DOI ↗