Towards the development of supervised machine learning algorithms to age-grade Aedes albopictus adults by mid-infrared spectroscopy

Mosquito population age structure is a crucial parameter for vector biology with relevant epidemiological implications and for the assessment of the effectiveness of control interventions. Nonetheless, conventional methods for estimating mosquito age are often slow, labor-intensive, and unreliable. However, Mid-Infrared Spectroscopy (MIRS) recently proved to be an efficient method to accurately assess the age of Anopheles malaria vector adults. Indeed, MIRS provides biochemical information of the mosquito cuticle based on the absorption of light, which changes during the ageing process. The complex information of MIRS spectra is then interpreted by Supervised Machine Learning (SML) algorithms to predict different age-classes. We investigated the potential of MIRS combined with SML algorithms to efficiently and accurately predict the age of adult females and males of Aedes albopictus, the main vector of dengue and Chikungunya in Europe. We analysed a total of 2,919 MIRS spectra from specimens reared under both laboratory and semi-field conditions and collected from emergence to 33 day old. The initial dataset consisted of 1,038 laboratory-reared individuals, demonstrating the feasibility of using spectroscopy for age classification. The algorithm was then refined using spectra from 1,881 semi-field mosquitoes reared under natural environmental conditions. In this case, we developed three algorithms with different resolutions: i) high resolution: age classes including individuals emerged within 3 consecutive days (1-3, 4-6 …); ii) medium resolution: classes of 6 days (1-6, 7-12 …); iii) low resolution: age classes of 9 days (1-9, 10-18 …. ). The highest age estimation accuracies were obtained with the low-resolution algorithm (≥83% for females and ≥98% for males), while the accuracies at the highest resolution were ≥52% and ≥81% respectively. The model at medium resolution obtained accuracies of ≥61% for females and ≥81% for males. Furthermore, we assessed the ability of our MIRS-SML algorithms to detect an age structure shift by simulating a control intervention on in-silico mosquito populations, comparing a control population with natural mortality (0.04%) to a population halved by a theoretical control intervention after one week. We showed that MIRS-SML can statistically detect a change in age structure just by measuring 100 samples of the target population. These findings underscore the high potential of this innovative approach to enhance the accuracy of epidemiological models aimed at estimating the risk of arbovirus transmission. Given the easiness to implement MIRS-SML, age-grading mosquito populations has the potential to become a complementary approach to the assessment of population abundance to quantify the effectiveness of vector control strategies. Further training of the ML algorithms with larger sample sizes maintained in more variable environmental conditions is needed in order to obtain higher predictive accuracies and to apply the approach in real life assessments.