Machine learning for syndromic surveillance using veterinary necropsy reports.

Area of Science:

• Veterinary Medicine
• Computational Biology
• Epidemiology

Background:

• Natural language data from animal health records offers potential for disease surveillance.
• Veterinary necropsy reports contain valuable, yet largely unstructured, information on animal diseases.
• Current surveillance methods may benefit from advanced data analysis techniques.

Purpose of the Study:

• To evaluate machine learning (ML) methods for syndromic surveillance using free-text veterinary necropsy reports.
• To develop and assess a system for detecting specific pathologies (gastrointestinal, respiratory, urinary) in necropsy reports.
• To apply a trained ML model to a large dataset for epidemiological trend analysis.

Main Methods:

• Trained ML algorithms, including deep learning (long short-term memory networks) and random forest, on veterinary necropsy reports.
• Utilized TF-IDF (term frequency-inverse document frequency) statistics for feature vector creation.
• Evaluated model performance using F1 scores for gastrointestinal, respiratory, and urinary pathology detection.

Main Results:

• Random forest models with TF-IDF features achieved high performance: F1 scores of 0.923 (gastrointestinal), 0.960 (respiratory), and 0.888 (urinary).
• The best-performing model was applied to over 33,000 necropsy reports spanning 14 years.
• The analysis revealed temporal and spatial disease patterns, including a potential gastrointestinal disease cluster in 2016.

Conclusions:

• Machine learning, particularly random forest with TF-IDF, is effective for syndromic surveillance using veterinary necropsy reports.
• This approach can uncover epidemiological trends and identify disease hotspots.
• Automated analysis of veterinary reports provides a valuable tool for monitoring animal health and detecting emerging threats.