Back to the Future: Quantifying Wing Wear as a Method to Measure Mosquito Age

ABSTRACT.

Vector biologists have long sought the ability to accurately quantify the age of wild mosquito populations, a metric used to measure vector control efficiency. This has proven challenging due to the difficulties of working in the field and the biological complexities of wild mosquitoes. Ideal age grading techniques must overcome both challenges while also providing epidemiologically relevant age measurements. Given these requirements, the Detinova parity technique, which estimates age from the mosquito ovary and tracheole skein morphology, has been most often used for mosquito age grading despite significant limitations, including being based solely on the physiology of ovarian development. Here, we have developed a modernized version of the original mosquito aging method that evaluated wing wear, expanding it to estimate mosquito chronological age from wing scale loss. We conducted laboratory experiments using adult Anopheles gambiae held in insectary cages or mesocosms, the latter of which also featured ivermectin bloodmeal treatments to change the population age structure. Mosquitoes were age graded by parity assessments and both human- and computational-based wing evaluations. Although the Detinova technique was not able to detect differences in age population structure between treated and control mesocosms, significant differences were apparent using the wing scale technique. Analysis of wing images using averaged left- and right-wing pixel intensity scores predicted mosquito age at high accuracy (overall test accuracy: 83.4%, average training accuracy: 89.7%). This suggests that this technique could be an accurate and practical tool for mosquito age grading though further evaluation in wild mosquito populations is required.

INTRODUCTION

An insect’s ability to act as a vector for human pathogens is quantitatively understood via vectorial capacity, a mathematical measurement of a vector’s ability to spread disease.¹ Among mosquitoes, females are the sole transmitters of pathogens to vertebrates as blood feeding is often required for egg development.^2,3 Most mosquito-borne pathogens ingested in bloodmeals develop and/or replicate in the midgut, and then migrate to the salivary glands or other specific tissues to be transmitted to the next host.^4–6 The time required for this process is known as the extrinsic incubation period (EIP), and it is often long relative to the average mosquito’s adult life span. Daily and lifetime survival probability are, therefore, the most influential determinants of vectorial capacity.^6,7 As such, to successfully prevent the transmission of mosquito-borne pathogens, control strategies are often designed to shorten mosquito life span below the EIP of a given pathogen. Therefore, it is useful for researchers and control programs to accurately measure epidemiologically meaningful changes in the age structure of mosquito populations.

Currently, the most widely accepted and used method for mosquito age classification is examining changes in ovary appearance. Although several techniques exist for assessing age based on mosquito reproductive morphology, the Detinova method remains the archetype due to its simplicity. Age assessment is conducted by characterizing the tracheole skeins that supply oxygen to individual ovarioles within each ovary, which remain tightly coiled for virgin females but permanently stretched and uncoiled after a gonotrophic cycle. As such, females can be age characterized as nulliparous or parous.^8,9 Despite this being used most often for age grading wild-type mosquitoes, it presents numerous challenges. First, it only provides an inexact binomial categorical classification of mosquito age based on parity status. Parous females, therefore, could be as young as 3–5 days or several weeks old.^2,3 Consequently, age classification by parity is limited in its ability to yield epidemiologically relevant information relative to a pathogen’s EIP. Second, the veracity of the method is questionable in mosquito species capable of gonotrophic discordance and autogeny.^9–11 In addition, parity assessment can be subjective in some species whereby ovarioles may develop unevenly during gonotrophic cycles and so a single ovary may contain tracheole skeins of mixed morphologies, making it difficult to classify parity status in an unbiased fashion.^9,12,13 Lastly, although the dissection techniques and tools needed for the Detinova method are relatively simple, dissecting ovaries from numerous mosquito samples on a regular basis in the field can be labor intensive. Alternative methods originally developed by Polovodova and modified by others build upon the Detinova method to calculate age from more advanced screenings of ovary morphology. These methods quantify the number of ovariole dilations and egg sac stages.^12,14–16 However, these techniques still measure age in terms of previous gonotrophic cycles. In addition, this technique is even more labor intensive than parity dissections, and can also be biased by inconsistent ovariole development in each ovary.

More modern tools for determining mosquito age include chemometrics,^7,17–21 transcriptional and protein profiling,^22–25 biochemical analyses,^26,27 and alternative physiology-based morphological assessments.^28–30 These methods present their own technical and cost–benefit challenges and all have biases in being able to accurately measure chronological age, especially from wild-type adults.³¹ An ideal mosquito age grading tool should be viable with both laboratory colonies and in field settings using wild-type adults, be neither technically difficult nor too labor intensive, and allow for rapid processing of many specimens per sampling time point. Similarly, it should be inexpensive, reliable, and accurate across mosquito species, and quantify mosquito age in a manner that is epidemiologically relevant.

The earliest assessments of mosquito age that we are aware of simply documented the external wear of mosquitoes, assuming wear increases with age due to the effects of frequent flying over time, and associated it with sporozoite rates. In 1912, Perry detailed his qualitative categorization of wild Myzomyia (Anopheles) spp. mosquitoes based on their wing condition. Calculating sporozoite rates from a subset of these wild-caught mosquitoes showed that all sporozoite-positive mosquitoes were classified as having grade 3 wing wear, which was described as being “decidedly shabby and the wing fringe very much worn.”³² This finding was supported nearly 20 years later by Gordon et al., who used the technique to age-grade wild anophelines in Sierra Leone and observed that sporozoite rates increased as wing condition deteriorated.³³ Despite the obvious correlation of wing condition and sporozoite rates, qualitative wing condition grading has been critiqued as being a subjective age-assessment tool and also biased by environmental conditions.^32,34 Recognizing these possible faults but also the potential ease of this assessment, we developed and evaluated quantitative metrics of Perry’s original technique. We hypothesized that wing scale loss correlates with time posteclosion due to normal flying activity, and thus could be used as a quantitative means of estimating mosquito chronological age. To test this hypothesis, we conducted laboratory experiments using Anopheles gambiae held as adults in either simple cages or more complex mesocosms that also featured ivermectin treatments. Mosquito age was then assessed by photographing the wings of mosquitoes under a microscope and either manually counting a subset of wing scales in each photograph or analyzing total wing wear through machine learning (ML) and pixel intensity algorithms that assessed the entire wing image.

MATERIALS AND METHODS

Mosquitoes.

Anopheles gambiae (G3 strain) were reared in an insectary at 27–30°C, 60–80% relative humidity, with a standard photoperiod (16 hours of light:8 hours of dark). Larvae were reared in open-faced plastic bins and fed a diet of ground TetraMin fish food (Spectum Brands Pet, LLC).³⁵ Pupae were transferred from larval rearing pans to closed, plastic emergence cages. Twenty-four hours posteclosion, adults were gently aspirated with a ProkoPack and moved to an adult-only colony bin.³⁶ Adults were binned according to age, with large cohorts of adults of the same age being placed together. In cases of smaller cohorts, adults of mixed ages were grouped together. However, these adults never differed by more than two days in age. Adults were fed 10% sucrose in water ad libitum.

Simple cage experiment.

For this experiment, parity status was classified per definitions set out by Benedict et al. Nulliparous status was defined not by whether a female mosquito had blood fed or not, but rather by her virgin status. Parous females were those who had oviposited at least once.³⁵

Approximately 300, 5- to 6-day-old female An. gambiae were transferred into a 30.5 × 35.5 × 40.5 cm plastic cage via ProkoPack aspiration. Mosquitoes were then fed defibrinated calf blood in a glass membrane feeder for 1 hour. After 30 minutes of rest, mosquitoes were knocked down with cold. Unfed females were removed and analyzed as unfed nulliparous. Blood-fed females and males were reintroduced into the cage. Three days later, 10 females were removed from the cage with a mouth aspirator and classified as blood-fed nulliparous. Mosquitoes were then given a second bloodmeal to allow for full ovary development. Afterward, they were provided damp paper towels for egg laying. Twenty-four hours later, the egg papers were removed, and 10 females were removed from the cage with a mouth aspirator and classified as uniparous (i.e., completion of one gonotrophic cycle). The remaining mosquitoes were given a third and final bloodmeal. Females that did not feed or fully engorge were removed via cold knock down. Engorged females and males were reintroduced and provided egg papers. Twenty-four hours later, egg papers were removed along with 10 biparous (i.e., completion of two gonotrophic cycles) females via mouth aspiration. The experiment was repeated three more times to have four replicates and 139 females sampled in total. In one replicate, enough mosquitoes remained after biparous sampling to go through a third blood feed and egg laying (triparous—completion of three gonotrophic cycles). Additionally, 20 teneral females were sampled < 24 hours after emergence to determine the base number of scales present along wing edges (described below).

Mesocosm experiments.

Two mesocosms (122 cm long × 61 cm wide × 96.5 cm height) were constructed using plastic containers (Figure 1). At the front of the mesocosm, mosquitoes had to pass through 2.5 cm holes cut into a Tupperware container (15 cm long × 15 cm wide × 15 cm height) to feed on sugar cubes. They had to enter a separate but identical container to lay eggs on damp paper. To obtain water and bloodmeals, mosquitoes had to traverse the length of the mesocosm, navigate through a tall, artificial plant, and enter a separate chamber in the back upper section of the cage. Initially, approximately 400, 3- to 5-day-old, male and non-blood-fed female An. gambiae were placed into each mesocosm. The age distribution between both mesocosms was approximately equal. Mosquitoes in both mesocosms were offered two defibrinated calf bloodmeals every week for 10 weeks. However, the intervention mesocosm received a bloodmeal containing 5 ng/µL ivermectin every other week. This concentration, which is less than the lethal concentration to kill 50% (LC₅₀) of this species, was previously shown to significantly reduce mosquito fecundity and affect survivorship but not induce mortality across the entire mosquito population.^37–39

Figure 1.

Mesocosm design and procedural overview for mesocosm-based experiments. Anopheles gambiae G3 mosquitoes were reared in two separate mesocosms for 10 weeks. Navigational flight was encouraged by placing (A) egg papers, (B) sugar, and (D) blood/water in different locations and chambers covered with small entry/exit holes. Mosquitoes were also forced to fly and experience wing stress by having to navigate through a (C) fake plant to reach blood and water. The control mesocosm received untreated blood, while the intervention mesocosm received a sublethal 5 ng/µL dose of ivermectin in its blood feed every other week. Every 2 weeks, 200 fluorescent powder-marked mosquitoes were added to each mesocosm; twice each week, (E) 30 unmarked mosquitoes were added to each mesocosm. Twice each week, (F) 15 females were removed from each mesocosm for parity and wing wear analyses. This figure appears in color at www.ajtmh.org.

The day before a blood feed, 30 unmarked mosquitoes (a mix of newly emerged males and females) were added to each mesocosm. To simulate a mass emergence event, every other week a normal input of 30 unmarked mosquitoes were replaced by a batch of 200 newly emerged, mixed male and female mosquitoes (Supplemental Figure 1). However, due to uncontrollable and unpredictable irregularities in mosquito emergence rates, there were some instances where an emergence group had < 200 adults. As such, some emergence groups had to be combined to have the required sample numbers. In these instances, mosquitoes added to each mesocosm were not all the same age but differed by no more than 1–2 days. When analyzing data from these mosquitoes, the average of the age groups was taken to have a single, numerical variable to use for analysis (i.e., if the group was 10–11 days old when sampled from a mesocosm, the average age was logged as 10.5). These mosquitoes were first collected from emergence bins with a ProkoPack aspirator, placed in a large Ziplock bag, knocked down with carbon dioxide, coated in a light dusting of fluorescent pigment powder (Shannon Luminous Materials, Inc.), and then transferred to a mesocosm. In total, five different colors were used, with each color used only once. As such, the exact age of a mosquito could be determined based on its color mark alone. The day after a blood feed, 15 females were removed from each mesocosm via mouth aspiration (total sampled: N = 537). Each group of removed females constituted a replicate. Females were selected in a random fashion, resulting in a mixed selection of unmarked mosquitoes of unknown ages and color-marked mosquitoes of known ages. Due to the long duration of the experiment (10 weeks) and the complexity of the continuously sampled and restocked mesocosms, only a single replicate was undertaken (one control and one treatment mesocosm).

Age assessments for both experiments.

Selected females were knocked down with cold and held on ice. For the simple cage experiments, these included newly emerged, unfed nulliparous, blood-fed nulliparous, uniparous, biparous, and triparous females. For the mesocosm experiments, 30 mosquitoes (15 from each mesocosm) were analyzed, and the color mark of each mosquito was also noted. Wings were carefully removed by first pinching the lateral sides of the scutum. This contracted the thoracic muscles, forcing the wings to flare away from the thorax at 90° or more. Wings were then removed by gently pulling the exposed wing at the basal junction to the thorax with a fine pair of forceps. Dissected wings were placed on a ∼5 µL drop of 10% sugar water on a glass slide, which was allowed to dry. The sugar water acted as an adherent, preventing the wings from shifting or losing scales when the slide was photographed at 30X with a Leica EZ4 W stereo microscope (Leica Camera AG). Wing photos were later analyzed by counting long wing scales present along the distal-posterior edge of the wing, as this is likely subject to the greatest rate of scale loss due to aerodynamic forces experienced during wing beats.⁴⁰ We only counted the tallest, long wing scales because they are the easiest to differentiate with the human eye. To further focus our effort, we only counted long wing scales within four wing cell regions defined by branches of the cubital (Cu) and medial (M) longitudinal veins that reached the wing margin: Cu1, Cu2, M2, and M4 (Figure 2).⁴¹ Ovaries were dissected from the same mosquitoes in water. Parity assessments were made according to the Detinova method with a Leica DM500 microscope (Leica Camera AG).⁸ These data are shown in Supplemental Table 1.

Figure 2.

Wing wear patterns and scale counting methods for age grading. Photos of female, lab-reared, G3 strain Anopheles gambiae wings at (A) 1 day, (B) 8 days, and (C) 14 days postemergence show that wing scales are lost as mosquitoes age. Wing scales vary in length along the (D) wing edge, (D and E) but only long wing scales were counted in four specific wing regions, Cu2, Cu1, M4, and M2 in our manual counting experiments. This figure appears in color at www.ajtmh.org.

Computational wing photo analyses.

In addition to manually counting wing scales in the predescribed regions, wing photos were also analyzed by computer algorithms to classify or measure total wing wear. First, ML techniques were used to classify adult female An. gambiae s.l. mosquitoes based on the malaria EIP. Although the EIP for malaria can depend on climatic conditions, it commonly ranges from 10 to 15 days for malaria.^42,43 To have the largest sample size possible for older class females, mosquitoes were categorized based on the lowest estimate for EIP (10 days old or older versus younger ages). All ML models were trained using PyTorch version 1.3.1 ML library in Python 3.6.10 version on an NVIDIA GeForce GTX 1060 GPU (NVIDIA Corp). There were 997 original images used but image augmentation using the computer programming library OpenCV (www.opencv.org; version 3.4.1) was implemented to increase the dataset size. This included image rotation, horizontal and vertical flipping, Gaussian blurring (5 × 5), and mean denoising. The initial ML classifier schema and test dataset did include all images, however, removing days 9–11 to distinguish mosquitoes more easily from young versus old groupings showed increased accuracy of the final algorithm when compared with training with all images. After image augmentation, there were 1,917 images for the > 10 days classification and 1,950 images for the ≤ 10 days classification. During the initial training, we explored combining both the cage and mesocosm data, but the final algorithm only included the mesocosm original and augmented images. The first ML algorithm techniques explored a variety of pretrained neural networks, which included ResNet-18, ResNet-50, ResNet-152, Densenet-121, and VGG-16.^44–46 The second approach explored was support vector machine (SVM) algorithms using the same PyTorch library.

Our last approach used traditional computer vision methodology, using OpenCV and NumPy to create a simpler and more robust classification model by computing the overall pixel intensities of wing images. With each wing photo (left and right for each mosquito), preprocessing in the form of histogram equalization was performed. Afterward, pixel intensity scores for each image were obtained by summing pixel intensities across the entire image, and these scores were subsequently recorded in Supplemental Table 1. Initially, a binary decision threshold model was created from the means of these pixel intensities to classify mosquitos ≥ 12 days old and ≤ 8 days old, with 9- 11-day-old mosquito samples being discarded.⁴⁷ This test showed that there was a marked difference between the old and young groups, and that using the pixel intensity method may even work for creating a linear regression model. Subsequently, the 9- to 11-day-old mosquito samples were reinserted into the dataset, and a linear regression model was generated for the entire dataset (control and ivermectin-treated mesocosms), with left- and right-wing pixel intensities averaged. Newly eclosed (day 1) mosquitos were additionally included. To test the reliability of our promising results, cross-validations were performed six times to evaluate the linear regression age grading model, with a training-test distribution of 80–20%. Results of this cross-validation are available in Supplemental Figure 2. Given the high accuracy of the cross-validation tests, a final linear regression model was used for age grading with the control and ivermectin-treated mesocosms.

Statistical analyses.

Any observation for which one wing region was not able to be counted (i.e., the wing folded on itself, scales were pulled away from the wing, photograph was slightly blurry, etc.) was removed from the dataset. For each remaining mosquito processed via wing scale counting, the values of the four wing regions (Cu2, Cu1, M4, and M2) were averaged over the left and right wings. Left and right wing counts were averaged when both values were available or taken as the value solely for the left or right wing otherwise. For wing scale counts from cage data, 139 mosquitoes sampled were included in the final analyses. Additionally, 519 mosquitoes were sampled from the mesocosm and included in the final analyses.

To determine whether mosquitoes displayed unequal rates of wing scale loss between left and right wings on any given individual, we conducted a basic correlation analysis comparing the total number of counted wing scales between the left and right wings for all mosquitoes. This was done for mosquitoes sampled from the cage experiments as well as the mesocosm. A univariate Poisson regression model was used to quantify the relationship between wing scale counts and age, by regressing age on the sum of all long wing scales across the Cu2, Cu1, M4, and M2 regions. For total wing wear quantification using wing photo pixel intensity, linear regression models were used to quantify the relationship between pixel intensity scores and age by interpolation. The proportions of young versus old mosquitoes of known and predicted ages in each mesocosm were analyzed using Fisher’s exact tests. To compare the initial rate of scale loss between the cage and mesocosm (control mesocosm only) experiment data, we used an F-test to compare the linear regression slopes between common early age points of the two experiments (day = 1 and day = 4.5). To determine whether there was a relationship between wing size and wing wear, we calculated the mean wing length for each control mesocosm mosquito by measuring this from the wing images using ImageJ (imagej.nih.gov/ij/index.html), and examined Pearson’s correlation and linear regressions between mean wing length and mean wing pixel intensity among the mosquitoes in each age grouping. All analyses were performed using SAS statistical software (SAS Institute; v. 9.4), GraphPad Prism (v. 8.1.1), and R (v. 6.3.1).

RESULTS

Age grading using wing scale counting from mosquitoes held in simple cages.

Parity assessments were made for a subset of mosquitoes before their first blood feed (nulliparous), while others were sampled post-blood feed and/or after a designated number of gonotrophic cycles. In all cases, the number of ovipositions and the age was known for all females. The mean age for nulliparous females was 5.6 days (N = 66; 95% CI: 5.4, 5.9) and for parous females was 12.7 days (N = 73; 95% CI: 12.0, 13.4), which were significantly different (t-test: t = 12.1, P value: < 0.0001).

Results indicated that there was a significant linear relationship between the total number of wing scales on the left (N = 133) and right (N = 135) wings (Pearson r: r = 0.65; P value: < 0.0001), demonstrating that scale loss rate was roughly equal for both wings. For the Poisson modeling analysis, a separate dataset from newly emerged, nulliparous females (N = 20) were added to the cage dataset to analyze chronological mosquito age ranging from 1 to 18 days postemergence from pupae. The results showed that counted long wing scales from each wing region decreased on average as mosquitoes aged (Figure 3), however, relatively more wing scales were lost from region Cu1 and relatively few were lost from wing region M2 as the mosquitoes aged. The trend of wing scale loss from aging was clearer when data were summed across all four wing regions (Figure 4).

Figure 3.

Wing scale loss in each counted wing region relative to mosquito age in simple cage experiments. Anopheles gambiae mosquitoes were sampled between 5 and 18 days postemergence. These time points corresponded with nulliparous and varying stages of parous ovary morphologies. Long wing scales were counted along the posterior, distal wing fringe within four individual wing regions: Cu2, Cu1, M4, and M2. Data from newly emerged females that were sampled from a separate cage at 1-day posteclosion are marked in red. Data from all known-aged females sampled from the cage are marked in black. Blue line denotes a Poisson regression model line discerned from the data (gray areas = 95% CI region of the model). This figure appears in color at www.ajtmh.org.

Figure 4.

Wing scale loss correlates with mosquito age in simple cage experiments. Anopheles gambiae mosquitoes were sampled between 5 and 18 days postemergence. Long wing scales were counted and summed across all four individual wing regions, Cu2, Cu1, M4, and M2. Data from newly emerged females that were sampled from a separate cage at one day posteclosion are marked in red. Data from all known-aged females sampled from the cage are marked in black. Blue line denotes the Poisson regression model line discerned from the data (gray areas = 95% CI region of the model). This figure appears in color at www.ajtmh.org.

Age grading using wing scale counting from mosquitoes held in mesocosms.

Poisson regression showed a similar relationship between wing scale loss and average mosquito age in the mesocosm experiments. Results indicated a significant decrease in total long wing scale counts over time in the control mesocosm (Pearson r: r = −0.47, P value: < 0.0001), but the initial rate of scale loss over time was significantly more pronounced in this experiment relative to the simple cage experiments (F = 32.2, P value < 0.0001; Figure 5).

Figure 5.

Wing scale loss correlates with mosquito age in mesocosm experiments. Anopheles gambiae mosquitoes were sampled twice per week from a mesocosm that was continuously replenished with marked mosquitoes over 10 weeks. The total number of wing scales was summed across the Cu2, Cu1, M4, and M2 regions and analyzed against the average age of marked mosquitoes measured in days. Data from newly emerged females that were sampled from a separate cage at 1-day posteclosion are marked in red. Data from all known-aged females sampled from the control mesocosm are marked in black. Blue line denotes the Poisson regression model line discerned from the data (gray areas = 95% CI region of the model). This figure appears in color at www.ajtmh.org.

Assessing ivermectin-induced effect on mesocosm age structure.

Parity analyses of females of known age showed significantly different mean ages between nulliparous and parous mosquitoes tested from the control mesocosm (nulliparous = 8.9 days [95% CI: 6.9, 11.0], N = 42; parous = 13.2 [95% CI: 12.2, 14.2], N = 115; t-test: t = 4.2, P value < 0.0001) and the intervention (ivermectin-treated) mesocosm (nulliparous = 7.4 days [95% CI: 5.7, 9.2], N = 32; parous = 11.5 [95% CI: 10.4, 12.6], N = 97; t-test: t = 3.7, P value = 0.0003). Furthermore, the mean ages of nulliparous and parous mosquitoes were lower in the intervention mesocosm relative to the control mesocosm. Although parity status proved to be an accurate predictor of chronological age for individual mosquitoes, accuracy decreased when populations of mosquitoes of mixed and unknown ages were analyzed. When analyzing all mosquitoes sampled over the course of the entire 10-week experiment, no significant difference in the proportion of parous mosquitoes could be detected when comparing the control and ivermectin-fed mesocosms (N = 519; χ² value: 0.26, df = 1, P value 0.61).

Pearson correlation coefficients indicated a statistically significant negative correlation between the two wings (r: r = −0.67, P value: < 0.001). This indicated that both wings could be treated equally. We next assessed whether wing scale counting could be used to detect the shifts in age structure induced through the ivermectin-containing blood feeds. We conducted two-way t-tests for wing scale counts in each individual wing region, as well as the total sum of wing scales for mosquitoes of both known and unknown ages grouped based on mesocosm. Whether summed across all wing regions or within an individual wing region, the average number of wing scales was always significantly higher for mosquitoes in the ivermectin-treated mesocosm than those in the control mesocosm (Table 1).

Table 1.

Two-way t-test for difference in wing scale counts from mosquitoes of known and unknown ages between both mesocosms

Wing region	Mesocosm	Mean n wing scales	95% CI	t-Test value	P value
Cu2	Control	14.1	(13.3, 14.8)	1.99	0.0477
	Ivermectin	15.1	(14.4, 15.8)
Cu1	Control	13.3	(12.5, 14.1)	2.61	0.0094
	Ivermectin	14.8	(14.0, 15.6)
M4	Control	7.7	(7.0, 8.3)	3.87	0.0001
	Ivermectin	9.5	(8.8, 10.1)
M2	Control	5.3	(4.7, 5.8)	2.98	0.0031
	Ivermectin	6.5	(5.9, 7.0)
All wing regions	Control	40.3	(37.8, 42.8)	3.17	0.0016
	Ivermectin	45.7	(43.4, 48.0)

Age grading by assessing total wing wear using computer vision algorithms.

Lastly, we investigated if computer vision algorithms could correctly classify or quantify mosquito age using wing photos derived from the mesocosm experiments. Machine learning outputs were first analyzed as a binary outcome of mosquitoes ≤ 10 days old (N = 235) and > 10 days old (N = 253). Early exploration utilizing mosquito wing photos from the simple cage experiment (non-mesocosm) using the ResNet18 convolutional neural network (CNN) showed high 93.7% accuracy on the test set. However, when using images generated from the more complex mesocosm experiment, the accuracy of the ML analysis using resnet18 CNN decreased to 84.2%. It was suspected that this was due to a greater diversification of wing scale states with the addition of environmental stressors in the mesocosm as compared with the cages. We then combined the images from the cage experiment and the mesocosm trial to increase the number of training set data points used in the model. Surprisingly, we saw a substantial decrease in accuracy to 75.7%. This indicates that the increased wing wear from the more complex mesocosm cohort expressed enough differences from the simpler cage experiments to support the need for environments that replicate wild-type ecosystems. To further explore where a natural maximum wing wear threshold occurred, we incrementally adjusted the classification cutoff to a lower time point with just the mesocosm dataset. We achieved the highest score of 88.0% accuracy when using the binary classification ≤ 6 days old and > 6 days old. This suggested that greater differences in the numbers of scales are seen at younger ages and that older aged mosquitoes are harder to differentiate with this method. This is further demonstrated by the nonlinear relationship between the number of scales and the true age observed in the regression analyses (Figure 5).

To increase the model’s accuracy in distinguishing mosquitoes older or younger than the malaria EIP (10 days), we used the combined data from the mesocosm and cage experiments but removed the image data that was taken from days 8 to 12. In contrast, our test set still contained the images from days 8 to 12. The pretrained CNN and SVM models achieved similar results (75.4% accuracy) as what was seen previously.

Image analysis using pixel intensity began by processing individual wing photos through algorithms to achieve histogram equalization. A linear binary decision threshold analysis was explored to categorize control mesocosm mosquitoes predicted to be above or below the EIP (10 days old). This model achieved an 88.0% accuracy between distinguishing individual mosquitoes of known age above and below this cutoff. In contrast, a human classifier counting only long wing scales in our predetermined wing fringe regions from the same mosquito wing photos correctly classified the age only 56.0% of the time.

In comparing the two mesocosms, the mean pixel intensity scores of all mosquitoes sampled were significantly different from each other (control = 119,162,337; ivermectin = 119,102,006; t-test: t = 2.8, df = 517, P value = 0.0061), reflecting a significantly younger age in the ivermectin mesocosm due to a lower pixel intensity score caused by darker pixels on wings with more scales. Subsequently, linear binary decision threshold analysis was applied to mosquitoes sampled from both mesocosm datasets to group them into young and old age classes. By averaging pixel intensity scores of ivermectin and control mesocosm mosquitoes of known age, a single, common threshold value of 119,141,236.5 was applied to the data. This threshold was then used to group mosquitoes of unknown ages into each age class (Figure 6).

Figure 6.

Mosquitoes from the control and ivermectin-treated mesocosms classified as young and old based on age and their mean wing image pixel intensity score. Orange dots represent mosquitoes of known age (marked) ≤ 8 days old, while orange squares represent unmarked mosquitoes predicted to be ≤ 8 days old. Blue dots represent mosquitoes of known age (marked) ≥ 8 days old, while blue squares represent unmarked mosquitoes predicted to be ≥ 12 days old. Mosquitoes graded as ages 9–11 were removed from the analysis for greater separation of the age groupings. The threshold was calculated by determining the mean pixel intensity of marked mosquitoes analyzed from both mesocosms. This figure appears in color at www.ajtmh.org.

Based on the results from the linear binary threshold, linear regression analysis was then performed. Using only images of wings from mosquitoes with known ages from both the control and ivermectin mesocosm cohorts, images were processed by calculating the left and right wing pixel intensity scores averaged per mosquito followed by cross-validation. The average overall test accuracy was 83.4% and the average training accuracy was 89.7%. The results of the six independent cross-validation trials are presented in Supplemental Figure 2. We also used these data to determine whether there were correlations between wing size and wing wear from control mosquitoes of each known age class, but saw no evidence of a relationship (Supplemental Figure 3).

Given the accuracy of the cross-validation tests using image pixel intensity, a summary analysis was made for all control and ivermectin-treated mesocosm mosquitoes of known age, as well as newly emerged mosquitoes (Figure 7). Overall, a higher proportion of mosquitoes from the ivermectin-treated mesocosm (108/257; 0.42 [95% CI: 0.36, 0.48]) were grouped into the younger age class relative to those in the control mesocosm (96/262; 0.37 [95% CI: 0.31, 0.43]), while a higher proportion of control mesocosm mosquitoes were grouped into the older age class (137/262; 0.52 [95% CI: 0.46, 0.58]) relative to the ivermectin-treated mesocosm (113/257; 0.44 [95% CI: 0.38, 0.50]). However, neither comparison was statistically significant.

Figure 7.

Linear regression of mean wing image pixel intensity per individual plotted by age. Each blue and orange dot represents a single marked mosquito. All color-marked mosquitoes from the control and the ivermectin mesocosm were included in the analysis. Image pixel intensity scores were derived from the mean score from left and right wings from an individual mosquito. Added green dots represent the newly-eclosed (day 1) mosquitoes. This figure appears in color at www.ajtmh.org.

DISCUSSION

Our data represent the critical first steps to developing a robust technique for age grading mosquitoes using wing wear. Assaying mosquitoes in laboratory-based cage experiments indicated that the number of long wing scales across the four predefined wing regions (Cu1, Cu2, M2, and M4) was reduced over time. This relationship was also observed in the mesocosm experiment, but the scale loss over time was more pronounced. While the wing scale counting technique was generally effective, the variance of scale loss per mosquito of known age was high and the process of manually counting scales was tedious and potentially biased. In contrast, computer vision techniques, particularly wing image pixel intensity analysis, showed great promise for grading mosquito chronological age due to its relatively high accuracy, the potential for automation of the process, and lower variance per mosquito of known age. Overall, sublethal exposure to ivermectin in one mesocosm led to significant shifts in the age structure toward younger classes of mosquitoes when compared with the control mesocosm that could be detected with both the wing scale counting technique and wing image pixel intensity analysis.

Mosquito age has long been recognized as a potentially good indicator of vector control efficiency and, more recently, is seen as a viable metric to predict mosquito-borne disease outbreaks.^48,49 However, because mosquito survivorship typically follows complex, age-dependent patterns,^50,51 new techniques are needed that can finely classify mosquitoes by chronological age. Currently, researchers and control personnel are limited in their predictive abilities as only a few age grading methods are able to calculate continuous age, notably gene and protein profiling, mark release recapture (MRR) studies, and chemometric methods (near-infrared spectroscopy [NIRS] and mid-infrared spectroscopy [MIRS]).³¹ Of these, NIRS has been the most widely used as NIRS scanning is rapid and field-viable, performs nondestructive analysis, and can also potentially detect Plasmodium infections and identify species and sub-species.^18,52 However, measurements of mosquito age from NIRS have comparatively high individual mosquito error.⁷ Difficulties are further augmented in the field, as NIRS data variation can be high likely due to the confounding effects of genetic, dietary, and environmental effects on wild mosquitoes.²⁰ In addition, instrument costs with NIRS-based analyses are also high, the statistical methods of analysis can be difficult for nonspecialists, and NIRS scan datasets cannot easily be combined across sites, often necessitating new calibration datasets for each mosquito population tested in time and space.⁷ Mid-infrared spectroscopy may allay some of these problems with NIRS. However, the technique is relatively new (and thus less tested) and also requires destructive measurements of the samples.³¹

Our results indicated that wing scale-based age estimation could prove to be a viable, alternative age grading technique. Even as early as 1912, Perry was able to successfully classify wild mosquitoes into four-tiered age groups based on the qualitative assessment of their wing.³² Our cage and mesocosm experiment results expand Perry’s original conclusions, showing that wing scale counting can use count-based data, and wing image pixel intensity can use continuous variable data, to calculate chronological age in days. This technique is also nondestructive, only requiring that wings be plucked from the sample during initial identifications under a microscope. Therefore, it allows for downstream processing of all other mosquito tissues. Both wing scale counting and wing image pixel intensity analyses were able to detect the changes in overall age structure between our two mesocosm populations due to natural flying activities of the mosquitoes that led to scale loss over time. The former technique showed relatively more scales in the select regions we counted from mosquitoes of younger age classes; the latter technique showed relatively lower (more black) mean pixel intensity scores of the entire wing images in younger age classes due to the presence of more dark scales on these mosquitoes. Overall, the control mesocosm’s population skewed toward older females, while the ivermectin-treated mesocosm skewed toward younger females.

Mosquito populations can be highly variable and change rapidly due to mass larval emergences or quickly changing environmental conditions at field sites. As such, many individuals should ideally be processed per time interval to achieve accurate measurement of the population age in that period. This also is often done by field teams who may have limited resources, familiarity with entomology fieldwork, or experience handling mosquito samples.⁵³ Considering these realities, an ideal age grading technique should be quick, low-cost, and be accessible even for novice personnel. Assessing wing wear via scale counting is inexpensive aside from the upfront purchase of a stereoscope with a camera. It is also intuitive and requires little training to master. To save time in the field, wing image analysis could be also done later from stored, slide-mounted wings, or the images could be immediately taken and virtually analyzed elsewhere. We have also demonstrated that workloads could be reduced further by reducing the number of wing regions used for counting; based on our goodness-of-fit statistics for Poisson modeling, counting scales in just the M2 and Cu2 regions predicted mosquito age just as accurately as counting long wing scales across all four regions (data not shown). However, focusing on only long wing scales on the distal posterior wing fringe may be biased by limiting the analysis to such a small part of the wing. To this point, we observed a high variance in scale counts per age on days 4–18. The variance may also be due partially to human error, as the counting process is tedious for the person who had to study many images. While mosquito age grading necessarily should be focused on the population rather than individuals, the two are related and it is significant that individual mosquitoes of known age were only accurately predicted to be above or below the EIP 56% of the time by the wing scale counting technique.⁷ We also observed a faster loss of scales in the mesocosm experiments that facilitated more mosquito flying relative to the simple cage experiments. This observation may be even more amplified between laboratory and wild environments, and the link between environmental conditions (e.g., excessive wind, temperature and humidity, physical encounters with available vegetation, the physical distance mosquitoes traverse between larval sites and households) and scale loss could confound comparisons of mosquito age across populations collected in time and space with this technique. Our calibration model might also be biased from the fluorescent powder marking we used. As such, it will be crucial for wing scale-based age grading to be performed and validated in field settings, and when using calibration models constructed with mosquitoes of known ages marked with different techniques and held in large outdoor mesocosms.⁵⁴ Regardless, some of these limitations in wing scale counting may be resolved by the image analysis approaches we tested.

Machine learning has become the gold standard in image analysis and has broadened the capabilities of scientific research, including in mosquito identification studies.^55,56 Here, we apply ML with image analysis to classify mosquito age. Using a variety of different algorithms, the application of ML to individual mosquitos’ wing images was easily able to outperform mosquito age classification by a human performing wing scale counting on the same images. While the advanced ML algorithms we tested (pretrained CNNs and SVM) were superior to human-based classification, the algorithm that showed the most promise, wing image pixel intensity quantification followed by simple linear regression, does not use computationally expensive ML algorithms but instead classifies mosquito age based on the intensities of the pixels located within the wing images. The advantage of this approach is that it (1) is intuitive (scales lost equate to higher/more white pixel intensity from the wing image), (2) requires far less computational resources, (3) is relatively easy to learn and perform in comparison to more technical ML-based image analysis techniques, and (4) can be used on standard computers without investing in costly graphics cards or processors. In contrast, the predictive variables that ML uses are generally unknown or hard to discern and it requires a large amount of data to be used in training, sometimes requiring thousands or more images to generate an accurate model.⁵⁷ Failure to collect the appropriate amount and quality of data may result in overfitting the model, increase its susceptibility to noise within the data, or otherwise decrease model performance. Because of the limited number of images available for training our ML classification models, all the concerns mentioned were issues we had to account for. Overall, we conclude that wing image pixel intensity analysis is the most advantageous method to classify the age of mosquitoes we have tested so far, but the more advanced ML models may prove better in time with the accumulation of more images from both the laboratory and field. In addition, collecting images in this manner allows for potential secondary analyses such as quantifying wing patterning to help in mosquito identification efforts. The pixel intensity data generally fit a linear regression but skewed most from this simple model at the later ages, at approximately 12 days old. Presumably, the wing images of the mosquitoes start to reach saturation of whiter pixel intensities as dark scale loss becomes very high in old age. Analyses of field mosquitoes with similar techniques in the future may be best analyzed with different regression models depending on scale color intensities of wild-type populations and the rate of wing scale loss in natural environments. Data could be analyzed on a continuous scale, binned into groups (e.g., 1–5, 6–10, > 10) or split into binary groups with a threshold on the most relevant epidemiological age of the mosquitoes, such as the EIP for Plasmodium development in An. gambiae.

Ovary assessments by the Detinova and Polovodova techniques have been the gold standard tool for entomologists in their effort to age-grade mosquitoes. Our results demonstrated that parity was reliable for determining an individual mosquito’s age in both our cage and mesocosm experiments. In samples where the true age of the mosquito could be externally verified, the average age for nulliparous and parous females aligned with what has been reported in the literature.^2,3 Yet, from an epidemiological standpoint, it is more productive to judge parity-based age grading in its ability to detect differences in the age structure of a mosquito population rather than of individuals. As evidenced in our mesocosm experiments, the Detinova method failed to detect the significant shifts in age structure induced by ivermectin in the mixed populations of known and unknown ages. In the RIMDAMAL clinical trial testing ivermectin for malaria control, no significant reduction in parity rates was observed between mosquitoes collected from control versus intervention sites. This result was perplexing, as indicators for reduced malaria transmission in ivermectin sites were observed, including participants’ exposure to biting anophelines.⁵⁸ However, sublethal ivermectin exposure is known to interrupt the mosquito reproductive cycle. As shown in a study with Aedes aegypti, ivermectin exposure in bloodmeals caused notable dysregulation in ovarian development, including slow blood digestion without the development of ovarian follicles, degeneration of primary follicles and formation of ovarian dilatations within 24 hours post-blood feed, a significant reduction in the rate of vitellogenesis and follicle development, and decreased egg production.⁵⁹ Additionally, field studies involving Anopheles feeding on ivermectin-treated cattle demonstrated that ivermectin exposure reduced mosquito egg production for up to 15 days.^39,60 It is reasonable to conclude that sublethal ivermectin exposures result in some parity misclassifications due to the dysregulated ovarian development. This could explain why parity rates in the RIMDAMAL trial did not match other entomological indicators for malaria control in the field, and why parity assessment alone was unable to detect age structure differences between our two mesocosms in this study. As such, having an accurate mosquito age grading technique that does not rely on ovarian development, such as NIRS, MIRS, and wing wear analysis, may be more appropriate for analyzing ivermectin-based control approaches. Similarly, other new and developing vector control tools should assess the interventions’ sublethal effects on mosquito ovarian development to avoid such potential confounders in the future.

In conclusion, assessing wing wear for mosquito age grading has promise to be both an accurate and practical tool to use in evaluating the risk of mosquito-borne pathogen transmission and the efficacy of mosquito-control interventions in research and operational settings. Further investigation is needed in other species, and whether wing wear in field settings, either done via human assessments or computational wing image analyses, is a germane alternative to other age grading techniques such as parity assessments and NIRS. Furthermore, it would be both insightful and important to determine in field settings if ovarian-based grading techniques remain a viable tool for vector control efficacy assessments when insecticides/endectocides that impair or delay female mosquito reproduction are used.

Note: Supplemental tables and figures appear at www.ajtmh.org.