Application of Large Language Models to Predict Enteric Viruses Removal Using Surrogate Indicators

Area of Science:

• Environmental microbiology
• Wastewater engineering
• Data science

Background:

• Wastewater enteric virus monitoring faces challenges due to limited, expensive data and unreliable surrogate markers.
• Accurate prediction of viral removal in water reclamation facilities is crucial for public health and effective wastewater surveillance.

Purpose of the Study:

• To introduce CALM-VR, a novel framework combining large language models and machine learning for enhanced enteric virus log removal prediction.
• To address data scarcity in wastewater surveillance by generating high-fidelity synthetic data.

Main Methods:

• Utilized GPT-5.2 for schema-constrained synthetic data generation, coupled with machine learning models.
• Employed 48 curated datasets from five US water reclamation facilities, including physicochemical, viral surrogate, and enteric virus data.
• Validated synthetic data fidelity through distributional agreement, correlation analysis, and principal component analysis.

Main Results:

• Synthetic data closely mirrored real data distributions and dependencies, with no significant post-correction differences.
• Machine learning models, particularly tree-based approaches, significantly outperformed benchmarks in predicting viral removal.
• Synthetic data augmentation dramatically improved prediction accuracy (R² ≈ 0.71–0.89) compared to using real data alone (R² ≈ 0.11–0.46).

Conclusions:

• CALM-VR offers a reproducible and quality-controlled method for improving enteric virus removal prediction in data-limited settings.
• The framework demonstrates the potential of AI-driven synthetic data generation to advance wastewater surveillance and public health protection.
• While cross-facility transferability requires further investigation, CALM-VR effectively enhances predictive capabilities under data scarcity.