Deep Learning Algorithms Improve Automated Identification of Chagas Disease Vectors

Area of Science:

• Entomology
• Public Health
• Computer Science

Background:

• Chagas disease, transmitted by triatomine vectors, significantly impacts public health in the Americas.
• Accurate identification of triatomine species is crucial for disease control but faces challenges due to limited entomological expertise.
• Previous work established an automated system for triatomine species identification from digital images.

Purpose of the Study:

• To enhance triatomine species classification accuracy using deep learning algorithms.
• To compare the performance of deep learning with statistical classifiers for triatomine identification.
• To assess the impact of incorporating distributional data on identification rates.

Main Methods:

• Employed TensorFlow, an open-source deep learning platform, to train a classification algorithm.
• Utilized datasets of 405 images for Mexican triatomine species and 1,584 images for Brazilian triatomine species.
• Integrated distributional information to refine species identification.

Main Results:

• The deep learning system achieved 83.0% accuracy for Mexican species and 86.7% for Brazilian species.
• This represents an improvement over statistical classifiers, which yielded 80.3% and 83.9% accuracy, respectively.
• Incorporating distributional data boosted identification rates to 95.8% for Mexican and 98.9% for Brazilian species.

Conclusions:

• Automated identification using deep learning offers a promising approach to address the 'taxonomic impediment' in Chagas disease vector control.
• The developed system demonstrates improved accuracy and efficiency in identifying triatomine species.
• Integrating AI with entomological knowledge can provide a partial solution to critical public health challenges posed by vector-borne diseases.