With the growing global threat of mosquito-borne diseases, there is an urgent need for faster, automated methods to assess disease load of mosquitoes and predict future outbreaks before they occur. Current surveillance practices rely on mosquito traps that require manual collection of samples over days or weeks, followed by labor-intensive lab testing methods like polymerase chain reaction (PCR) and enzyme-linked immunosorbent assay (ELISA). These processes are time-consuming and resource-intensive, highlighting the need for innovative solutions that deliver rapid, real-time insights into vector infection status. In this study, we applied various machine learning algorithms to detect dengue-infected and noninfected mosquitoes based on their three-dimensional flight patterns. Our methodology involved using a convolutional neural network (CNN) and cubic spline interpolation to detect and track mosquito flight trajectories, followed by classification using machine learning algorithms, including CNN, XGBoost, AdaBoost, Random Forest, Decision Tree, Naive Bayes, Logistic Regression, Multi-Layer Perceptron (MLP), and a hybrid CNN + XGBoost model. Depending on sequence size, up to 43,278 three-dimensional flight trajectory sequences were used for dengue-infected and noninfected mosquito groups. Based on the mean values of 5-fold cross-validation, the results showed that XGBoost achieved the highest accuracy of 81.43%, closely followed by AdaBoost with 81.31% and Random Forest with 81.12%. In terms of F1 Score, Random Forest exhibited the best performance at 82.80%, while AdaBoost and XGBoost followed with F1 Scores of 82.44% and 82.22%, respectively. Across some folds, the models achieved outstanding performance. For example, in Fold 1, AdaBoost reached 95.85% accuracy with an F1 Score of 95.93%, while Random Forest achieved a recall of 97.77%. The study also analyzed the impact of flight sequence size on models' performance by varying sequence sizes between 50 and 250. Results indicated a direct relationship between sequence size and model performance, with longer sequences providing more accurate predictions. This study demonstrates the potential of artificial intelligence-driven models to enhance mosquito surveillance by automating the detection of infected mosquitoes. By providing a faster and more efficient method for assessing infection status, this approach can support real-time vector monitoring, improve early detection of disease outbreaks, and ultimately contribute to more effective mosquito control strategies.
