Enhancing decision confidence in AI using Monte Carlo dropout for Raman spectra classification

Area of Science:

• Microbiology
• Machine Learning
• Spectroscopy

Background:

• Machine learning (ML) models, including convolutional neural networks (CNNs), are standard for bacterial strain identification via Raman spectroscopy.
• Typically, ML models are trained and then used for inference without further enhancement, assuming peak performance post-training.
• This study addresses the critical, often overlooked, aspect of model uncertainty in deep learning.

Purpose of the Study:

• To develop and validate a novel methodology that quantifies and utilizes model uncertainty during the inference phase for improved bacterial identification.
• To enhance the reliability and accuracy of ML-based microbial identification by focusing on predictions with higher confidence.

Main Methods:

• Combined Monte Carlo Dropout (MCD) with CNNs, enabling uncertainty measurement during inference.
• Utilized Gaussian Mixture Model (GMM) to define an uncertainty threshold for categorizing unseen data.
• Focused final predictions on data subsets exhibiting lower model uncertainty.

Main Results:

• Applied the uncertainty-guided method to two Raman spectra datasets, showing significant accuracy improvements.
• Dataset 1 accuracy increased by 9% (83.10% to 92.10%) on a subset of 826 spectra.
• Dataset 2 accuracy increased by 12.82% (83.86% to 96.68%) on a subset of 1700 spectra.

Conclusions:

• Uncertainty-guided prediction is more effective for ensuring high prediction rates than using the entire dataset.
• This methodology enhances classification accuracy in critical applications like disease diagnosis and safety monitoring.
• The approach advances microbial identification, yielding more trustworthy predictions.