Novel Deep-Learning Unsupervised Domain Adaptation Method for Mitigating Batch, Strain, and Instrument Variations to...

Area of Science:

• Spectroscopy
• Machine Learning
• Microbiology

Background:

• Antibiotic resistance necessitates rapid pathogen identification.
• Raman spectroscopy with deep learning offers fast detection but suffers from domain shifts due to instrument and sample variability.
• Existing methods lack robustness against diverse spectral data.

Purpose of the Study:

• To develop a novel domain adaptation framework, the Raman Spectral Classification Discrepancy Model (RSCDM).
• To enhance the robustness and accuracy of bacterial identification using Raman spectroscopy.
• To address challenges in spectral domain adaptation for clinical applications.

Main Methods:

• Proposed the Raman Spectral Classification Discrepancy Model (RSCDM) for dynamic out-of-distribution sample detection.
• Employed adversarial feature alignment to bridge spectral variability across different domains.
• Utilized classifier output discrepancies for adaptive, task-driven domain alignment.

Main Results:

• RSCDM improved bacterial classification accuracy from 81.6% to 95.4% on a commercial spectrometer and 77.5% to 91.3% on a home-built spectrometer.
• Fine-tuning further boosted clinical isolate identification accuracy to 99.3%, demonstrating robustness to instrument diversity.
• Reduced bacterial identification time from days to minutes under high-load conditions.

Conclusions:

• RSCDM effectively addresses critical challenges in spectral domain adaptation for Raman spectroscopy.
• The framework enhances robustness to batch, strain, and instrument variability.
• Findings support the clinical translation of single-cell Raman spectroscopy for rapid pathogen identification.