The MosAICo ecosystem: bridging the taxonomic gap in vector surveillance with real-time entomological artificial intelligence

Technical Summary: The MosAICo Ecosystem

Problem Statement
Mosquito-borne diseases are escalating as a global health threat due to climate change, urbanization, and the expansion of invasive vectors. Current surveillance efforts face a critical bottleneck known as the "taxonomic impediment": the volume of mosquito specimens collected in the field vastly exceeds the capacity of expert entomologists to process and identify them. This disparity limits the effectiveness of real-time vector surveillance and hinders rapid public health responses.

Methodology
To address this bottleneck, the authors developed MosAICo, an integrated ecosystem combining standardized hardware and artificial intelligence for automated mosquito species identification. The system comprises two primary components:

1. Hardware: A standardized benchtop imaging device designed for high-throughput capture of trap catches.
2. Software (MosAICo-Net): A deep learning pipeline engineered for efficient, principled open-set recognition and uncertainty quantification. Unlike closed-set classifiers that force all inputs into known categories, MosAICo-Net is designed to identify target species while explicitly rejecting out-of-distribution (OOD) specimens that do not match the training distribution.

The model was trained and evaluated on a dataset of 12,499 specimens representing 15 species collected across Italy. It focuses on identifying seven priority vector species. The system's performance was validated through three distinct phases:

• Geographically Stratified Testing: Evaluation on a held-out test set to assess generalization within the training region.
• Field Validation: Deployment across 20 Italian surveillance sites involving 1,470 field-collected target specimens, with results compared against expert manual counts.
• Cross-Geographic Generalizability: Evaluation on 118 Aedes albopictus specimens collected at the invasion front in Ghana to test performance in distant, epidemiologically critical regions.

Key Results

• Target Species Identification: On the geographically stratified held-out test set, MosAICo-Net achieved >90% accuracy on target species.
• Out-of-Distribution Detection: The system demonstrated an AUROC of 0.96 for detecting specimens outside the training distribution.
• Field Performance: In the 20-site field validation, the system achieved 94% micro accuracy on target specimens and showed strong agreement with expert manual counts (Chi-square = 0.66).
• Cross-Geographic Robustness: When tested on Ghanaian specimens, the system achieved 97.4% accuracy, with only a single specimen requiring escalation to expert review.
• Throughput: The system processes up to 82 specimens per image, matching expert throughput speeds regardless of taxonomic complexity.

Significance and Claims
The paper positions MosAICo as a scalable "force multiplier" for public health entomology. By embedding uncertainty-aware AI within a standardized hardware-software pipeline, the system aims to resolve the taxonomic impediment by automating the identification of common vector species. The authors claim that this approach frees expert attention for the specific cases that require human validation: rare, invasive, or ambiguous specimens. The successful cross-geographic validation suggests the system is well-suited for deployment in distant regions where expert entomological capacity may be limited, thereby bridging the gap between specimen collection rates and identification capacity.