Sharpness-Aware Minimization (SAM) Improves Classification Accuracy of Bacterial Raman Spectral Data Enabling Portable Diagnostics

Abstract

Antimicrobial resistance is expected to claim 10 million lives per year by 2050, and resource-limited regions are most affected. In these regions, access to quality hospital-level clinical care and microbiological testing is limited. Thus, it is important to develop a solution for bacterial infection diagnosis that is rapid, reliable, and compact. Two techniques that can help contribute to more reliable diagnostic tools are Raman spectroscopy and deep learning (DL). Raman spectroscopy is a novel pathogen diagnostic approach that promises rapid and portable antibiotic resistance, and can provide results significantly faster then culturing methods. Despite its advantages, current algorithms for Raman spectral analysis 1) are unable to generalize well on limited datasets across diverse patient populations; and 2) require increased complexity due to the low-quality nature of Raman spectral data. In this work, we address the intrinsic challenges of working with Raman spectral data by utilizing Sharpness-Aware Minimization-based (SAM) optimizers for Raman spectral data analysis. We specifically observe this in clinical and non-clinical bacterial isolate classification tasks. We demonstrate that SAM-based optimizers achieve accuracy improvements on bacterial classification tasks. These results display the capability of SAM to advance the application of AI-powered Raman spectroscopy tools, enabling progress towards the translation of rapid on-device pathogen diagnostics for patients in need.