Exploring near-infrared spectroscopy ability to predict the age and species of Anopheles gambiae sensu lato mosquitoes from different environmental conditions in Burkina Faso

Abstract

Background

Near infrared spectroscopy (NIRS) has shown ability in previous studies to predict age and species of laboratory-reared and wild mosquitoes with moderate to high accuracy. To validate the technique as a routine tool, it is necessary to assess NIRS accuracy on these variables under different environmental conditions susceptible to affect the mosquito cuticle and interfere with the machine accuracy. This study investigated the influence of environmental conditions on NIRS accuracy to determine the age and species of Anopheles gambiae sensu lato (s.l.).

Methods

Environmental conditions of three important seasonal periods in Burkina Faso covering the onset, the peak and the end of the rainy season were mimicked in the laboratory using incubators. Emerged An. gambiae s.s. and An. coluzzii from laboratory colonies were reared in each period using temperature and relative humidity for predicting mosquito species by NIRS. Wild An. gambiae s.l. (n = 3788) were caught during the 3 different periods described above and analysed by NIRS to predict Anopheles species. Furthermore, first generation of wild Anopheles (n = 1014) was used to assess NIRS ability to classify mosquito age in each environmental condition. All data analysis were performed using a binomial logistic regression model.

Results

NIRS discriminated between laboratory-reared Anopheles with 83% of accuracy independently of any environmental condition. Similar trend was found in wild-caught Anopheles. NIRS accuracies varied slightly in laboratory Anopheles (77–85%) and more strongly in their field counterparts (67–84%). In both cases, models developed from the season of interest were more accurate than models trained with insectary conditions or from a different period of the year, indicating temperature and humidity can impact NIRS accuracy. Models derived from laboratory-mosquitoes reared under fluctuating environmental conditions predicted field-derived mosquito species with a low accuracy (59%). Models trained on varying conditions reliably classified age into two categories (< 9 days or ≥ 9 days, 79–84% accuracy).

Conclusion

NIRS was able to predict An. gambiae s.l. species and classified age into two categories under different environmental conditions with modest accuracy. Models trained using wild mosquitoes from one season could predict species in wild mosquitoes from a different season, though with slightly lower accuracy.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12936-025-05554-9.

Background

Malaria is the most widespread parasitic disease transmitted by Anopheles mosquitoes. Insecticide-treated nets (ITNs) and indoor residual spraying (IRS) have been the most important disease control interventions in endemic countries in Africa [1]. Vector control interventions aim to reduce Anopheles abundance or/and reduce their lifespan to limit Plasmodium transmission [2]. Entomological monitoring is needed to estimate the ability of the mosquito population to transmit the disease and evaluate control interventions [2]. These measurements would benefit from an understanding of the age and species of Anopheles vector populations. Current methods of estimating these entomological parameters are limited [3, 4]. For instance, molecular analysis methods to determine Anopheles species are technically laborious, require reagents, time-consuming and qualified workers, limiting the number of mosquitoes which can be processed. Also, age determination techniques based on observations of morphological changes in the reproductive system of female mosquitoes are laborious and require skilled workers, reducing their use on a large scale. To minimize these shortcomings, a variety of complementary or alternative methods are under development [3, 5–7]. NIRS, an innovative tool has the potential in predicting the entomological parameters of malaria such as age and species. This method has successfully identified laboratory and field Anopheles species [3, 8], age-grade [3, 7–10] and detected mosquito Plasmodium infection status [11, 12]. Compared to other methods, NIRS is fast, non-destructive, reagent free, waste free and requires less qualified personnel [4, 6, 11, 12]. While the technique appears promising, there are some challenges that must be addressed before NIRS can be used as a routine tool in the field [13]. For instance, the machines learning models need to be expanded using samples from a much wider diversity of mosquitoes, as models trained with mosquito data from one Anopheles population fail to reliably predict the characteristic of interest in other populations [10]. Models trained with spectra generated from laboratory-reared mosquitoes have been shown to predict field-mosquitoes age with relatively low accuracy [12]. Other studies have reported that the physiological status [14] and diet [15, 16] affect NIRS ability to accurately predict mosquito age. NIRS relies on the property of hydrocarbons within the mosquito cuticle to quantitatively measure changes in organic compounds such as C–H, O–H, N–H, and C–O functional. It detects these changes through the rotation, bending and stretching of these functional groups following exposure to light of the near infrared spectrum. The lipid structure of the arthropod cuticle depends on their environmental conditions [17], leading to NIRS spectral variations.

Vector control programmes collect mosquitoes in different locations at different periods of the year. Each of these contexts is characterized by particular environmental conditions such as temperature, relative humidity, photoperiodicity and ecological factors which may be able to induce morphological, physiological and biochemical changes within the mosquitoes. Recent work has reported that An. coluzzii and An. gambiae s.l. grow faster and live shorter at the peak of the rainy season (August) than their counterparts at the onset of the dry season (December) [18]. Furthermore, mosquitoes of An. gambiae complex reared under conditions typical of the onset of the dry season (December–February) have a larger body size than those reared under conditions from the late dry season (April–May) and those of the peak of the rainy season (August) in the Sahel region [19, 20]. It is thought that An. gambiae reared under dry season conditions have a higher hydrocarbon content compared to rainy season specimens [20]. These same authors reported higher chain lengths in n-alkanes in response to the dry season. These findings suggest seasonality induced cuticular hydrocarbons reorganizations within mosquitoes which could potentially influence NIRS accuracy. This could be a challenge for predicting sample characteristics for a given condition on the basis of models derived from another environmental condition. Verifying the dependence of NIRS predictions to environmental conditions is an important step for implementing NIRS technique as a routine diagnostic tool. This study aims to explore NIRS ability to predict species and age-grade of An. gambiae s.l. mosquitoes under different temperature and humidity in Burkina Faso at three time periods (the onset, peak and end of the rainy season) and compare these to mosquitoes reared under insectary conditions.

Methods

Experimental design

This study evaluated NIRS accuracy to identify species and age-grade of An. gambiae s.l. population reared under different environmental conditions by mimicking the important malaria seasons observed in Burkina Faso. In this country, entomological variations (vector density, mosquito longevity) occur at three different periods of the year [21]. June period is the onset of the rainy season characterized by significant daily fluctuations in air temperature and a considerable drop in relative humidity. During this period the vectors return from aestivation and the malaria transmission repeats. The peak of the rainy season (August) is the second important period associated with low variations of temperature and relative humidity. This period is characterized by a high density of mosquitoes associated with a high transmission of Plasmodium. The third important period is the end of rainy season/onset of the dry season (December) during which important daily fluctuations in air temperature and a significant drop in relative humidity were observed, marks the mosquito aestivation period [22]. Wild Anopheles were sampled at different time points of the year (onset, peak and end of rainy season or onset of dry season) for age determination or species identification using the NIRS technique. Anopheles coluzzii and An. gambiae s.s., two sibling species of the Gambiae complex, used in this study, significantly contribute to malaria transmission in Burkina Faso [23]. In the laboratory, these species were separately reared, while the field ones were identified using conventional PCR [26] after analysis with NIRS.

Environmental parameters simulation inside the incubator

In the IRSS laboratory, three incubators (Sanyo MLR 315H, Sanyo Electric Co., Osaka, Japan) were used to mimic cyclically environmental conditions of June, August and December for mosquito rearing [19, 22]. For each period, average hourly temperature and relative humidity (RH) (Fig. 1) cycles were scheduled inside the climate chambers based on data recorded hourly in the village of Bama in previous work [19]. To mimic natural daily climate fluctuations, a 12-step cycle of temperature and RH was set up in the climate chambers. A photoperiod of light/dark 12:12 h cycle in all climate chambers was performed for all experiments due to the low variations in photoperiod in the tropics (about 1 h 15 min day length variation between the solstices in Bobo-Dioulasso, Burkina-Faso). In another incubator, insectary conditions were programmed (temperature 27 °C ± 2, relative humidity 70% ± 5 with light/dark 12:12 h cycle) and acted as a control.

Fig. 1.

Environmental parameters of temperature (red line) and relative humidity (blue line) variation of the different periods of the year. The hourly mean variation of temperature and relative humidity recorded during the months of June (onset of the rainy season, at the left), August (peak of the rainy season, at the middle) and December (onset of the dry season, at the right) in the village of Bama (western Burkina Faso). The values recorded, were reproduced in the incubators to simulate these environmental conditions

Mosquito rearing inside the incubator for NIRS analysis

Species identification of laboratory An. gambiae s.s. and An. coluzzii

In the first part of the experiment, the effect of environmental changes on NIRS accuracy using adult mosquitoes was tested. Laboratory-reared female An. gambiae s.s. and An. coluzzii were sourced from an outbred colony established in 2018. Since changes in mosquito cuticles happen during adulthood [20], larvae were developed under controlled conditions (temperature 27 °C ± 2, relative humidity 70% ± 5 with light/dark 12:12 h) and then the resulting adult females were exposed under four environmental conditions. This approach could allow changes in the cuticular structure of mosquitoes exposed to different environmental conditions. Briefly, eggs laid by females of An. gambiae s.s. and An. coluzzii were separately hatched in trays containing water in the insectary. The resulting larvae were fed ad libitum with fish food (Tetramin^® Baby) and pupae were pipetted into a plastic cup half-filled with spring water and then placed in cages covered with untreated mosquito netting for emergence. Emerged female mosquitoes were organized by groups in cages and then kept in incubators where environmental conditions were programmed as designed below (Fig. 2). Mosquitoes were fed ad libitum with 10% glucose solution that was changed daily to avoid mosquitoes contaminated by mould. For a given generation of adult’s mosquitoes, approximatively 60 females of An. gambiae s.s. and An. coluzzii in each experimental group were sampled at different time point (3, 6, 9, 12, 15 and 18 days post-emergence) for NIRS analysis to identify Anopheles species. To minimize any interference effect and increase sample size, the experiment was repeated across four consecutive mosquito generations, resulting in four replicates. To avoid any effect of the cage, the positions were changed over time in climate chambers. In addition, climate chambers were also changed between conditions to avoid any interference of the incubator in data analysis.

Fig. 2.

Experimental design and sample size of Anopheles per environmental condition. N indicates the number of survival mosquitoes which were scanned by NIRS. These mosquitoes were used to explore mosquito species

Species identification of field-derived An. gambiae s.l.

The influence of different environmental conditions on NIRS accuracy in predicting Anopheles species in the field was evaluated and compared with laboratory-reared sample predictions. Wild mosquitoes were collected in the villages surrounding Bobo-Dioulasso at the periods with similar conditions which were reproduced in the incubators: June, July–September, October–November corresponding respectively to the onset, the peak and the end of rainy season. For each period, mosquitoes were caught early in the morning in the living rooms dwellings using mouth aspirators [24]. Anopheles gambiae s.l. morphologically identified based on Giles and Coetzee keys [25], were transferred at IRSS insectary in standard conditions (temperature 27 ± 2 °C, 70 ± 5% relative humidity, 12:12 light/dark). To prevent interference from blood meals in the NIRS spectra, unfed Anopheles mosquitoes were immediately scanned using the machine. After each mosquito scanning, the carcass was analysed for species determination using conventional PCR protocol [26].

Wild-mosquito age determination

First generation (F₁) of wild female An. gambiae s.l. were used in this experiment. Blood-engorged An. gambiae s.l. were caught during rainy season early in the morning using a mouth aspirator [24] in the living rooms of human in the villages of Bama and Soumousso. As reported in previous studies [27, 28], almost 99% of Bama’s mosquitoes are An. coluzzii while those from Soumousso include three major species (An. gambiae s.s., An. arabiensis and An. coluzzii). Collected mosquitoes were kept live in cages according to the locality, fed with 10% glucose solution during couple days for eggs laying as described in previous work [18]. Mosquitoes were killed after oviposition, individually preserved in Eppendorf tubes and then analysed by conventional PCR for species identification [26]. Eggs collected from F₀ females of each locality, were organized by pools and then incubated in different conditions (insectary, August and December) to obtain F₁ generation mosquitoes. The adult female mosquitoes were kept in the same conditions, fed with 10% glucose sugar solution and, around 30 Anopheles were sampled every 3 days (from day 3 to 18 day or day 21 according to environmental conditions) for NIRS analysis to determine Anopheles age. Two replicates were performed for this experiment.

Mosquito scanning

Mosquitoes from each experiment were killed with chloroform vapor and their cephalo-thorax scanned using a LabSpec4 Standard-Res i (standard resolution, integrated light source) near-infrared spectrometer and a bifurcated reflectance probe mounted 2 mm from a spectral on white reference panel (ASD Inc.). RS3 spectral acquisition software (ASD Inc.) which averages spectra from 20 scans, was used to record absorbance at 2151 wavelengths from 350 to 2500 nm of the electromagnetic spectrum. All specimens were scanned on the two sides focusing the light probe on the head and thorax region.

Data analysis

Data analyses were conducted using R software version 4.0.2. The wavelengths from 700 to 2350 nm were used to remove the excess noise arising from the sensitivity of the spectrometer often caused by too much light in the visible region. PLS (partial least squares) regression was used to analyse the relationship between spectra and entomological parameters (age and species) using packages downloaded from https://github.com/pmesperanca/mlevcm [29]. The age of each Anopheles was predicted using its spectrum. Additionally, a machine learning approach was used to develop binomial logistic regression model for performing Anopheles species identification and age classification (< 9 days or ≥ 9 days) [12, 29]. In this analysis, two response classes y = 0 and y = 1 were attributed to An. gambiae s.s. and An. coluzzii, respectively. The same procedure was used when classifying ages into two groups: y < 9 days for young mosquitoes and y ≥ 9 days for old mosquitoes. All analyses were performed using simple models without spectra smoothing or penalized estimation of the coefficient function. For reducing the overfitting, data were split into three subsets: 60%, 20% and 20% for training, validation and testing (estimation of the generalized error) the model, respectively. Sixty percent (60%) of sample was used to fit the model to known (age or species) sample using different numbers of PLS components. The first twenty percent (20%) corresponded to the sample used to select an optimum number of PLS component for the model. The second twenty percent (20%) of sample was used to assess the final model against a blinded subset of data. The number of principal components was allowed to vary between 2 and 50. This process was repeated 50 times to reduce errors. The observations were balanced by random sampling of the total number of mosquitoes analysed in all models. Model accuracy was obtained through the value of the area under the receiver operating characteristic (ROC) curve (AUC). A value of AUC close to or equal 1 indicates a good accuracy. The accuracies from different models were compared using Fisher exact test. Since mosquitoes were reared in different conditions, a calibration model developed using samples from one condition was used to predict the species or age of mosquitoes from another condition.

Results

NIRS ability to identify An. gambiae s.l. species from different environmental conditions

A total of 4131 laboratory derived Anopheles of various ages and reared under different environmental conditions were used to assess NIRS accuracy. In the field, 4343 mosquitoes were collected, with 12.8%, 41.5%, and 45.7% of An. arabiensis, An. gambiae s.s. and An. coluzzii respectively. Due to the low sample size of An. arabiensis, only spectra derived from An. gambiae s.s. and An. coluzzii were included in the data analysis. The preliminary analyses revealed distinct spectral profiles between the two groups of mosquitoes (laboratory-mosquitoes versus field mosquitoes) (Fig. 3).

Fig. 3.

Average spectra of laboratory-derived mosquitoes (red curve) and field-derived mosquitoes (blue curve); interval of wavelengths [700, 2350]

The binomial logistic classification model trained on mosquitoes reared across all conditions in laboratory discriminated between An. gambiae s.s. and An. coluzzii with 83% of accuracy (An. gambiae s.s. = 84%; An. coluzzii = 82%; Fig. 4a). When a calibration model was generated from laboratory mosquitoes reared in one condition to predict those in other conditions, NIRS accuracies varied from 77 to 85% (Fig. 5a). In most situations, models trained with mosquitoes from the same condition were more accurate, i.e. models calibrated with mosquitoes reared under August conditions were the most reliable in predicting mosquito species reared under August conditions. The lowest accuracy was obtained when models derived from Anopheles reared under June conditions were used to predict Anopheles species from other rearing conditions. However, these models were also trained on the smallest number of mosquitoes. In contrast, the high accuracy was found in August conditions when the model derived from mosquitoes in these conditions were used to predict Anopheles species from another reared conditions (Fig. 5a). The overall accuracies of different comparisons using a predictive model from one condition to identify mosquito species across all conditions resulted no difference for distinguishing An. gambiae s.l. species (p = 1). Altogether, models developed from insectary-reared mosquitoes and those from the peak of the rainy season were suitable for predicting species under different environmental conditions (Fig. 5a). Details of Anopheles species classification accuracy (An. gambiae s.s. and An. coluzzii) under each environmental condition were shown in Additional file 1: Table S1.

Fig. 4.

NIRS accuracy to predict mosquito species reared under all conditions. a Laboratory derived Anopheles; model trained using mosquitoes from all environmental conditions and b Wild caught Anopheles; model trained with mosquitoes from all collection periods

Fig. 5.

NIRS accuracy to predict An. gambiae s.s. and An. coluzzii species using calibration models from each environmental condition. a Laboratory derived Anopheles; b Wild derived Anopheles. Predictive models were trained on sample derived from one environmental condition to predict those from another condition

The analysis was extended to investigate whether these finding could be generalized to wild Anopheles. The model developed using wild mosquitoes across all periods classified An. gambiae s.s. and An. coluzzii with 84% accuracy (An. gambiae s.s. 86%; An. coluzzii 82%, Fig. 4b). This accuracy was not significantly different from that obtained with laboratory-reared Anopheles (84% versus 83%; p = 0.23). Again, accuracies varied considerably between 67 and 84% when the model trained with samples collected in one period was used to predict samples from another period (Fig. 5b). Indeed, the model derived from mosquitoes at the end of the rainy season generated low accuracy (67%) in predicting onset-season mosquito species, though these models were trained on the smallest number of mosquitoes. The highest accuracy (84%) was found with the model trained on samples from the peak of the rainy season. Additional file 2: Table S2 provided detailed classification accuracy for An. gambiae s.s. and An. coluzzii under each collected period. As in laboratory Anopheles, the model trained from the peak of the rainy season specimens had the highest accuracy across all populations (Fig. 5b). Finally, models derived from laboratory-mosquitoes reared under fluctuating environmental conditions predicted field-derived mosquito species with a low accuracy of 59% (An. gambiae s.s. 59%; An. coluzzii 58%).

NIRS accuracy to predict wild An. gambiae s.l. age under different environmental conditions

Around 2500 wild Anopheles were collected in the field, with 1000 females from Bama and 1500 from Soumousso. As expected, the three major species of Anopheles involved in malaria transmission (51.6% An. gambiae s.s., 38.2% An. arabiensis and 10.2% An. coluzzii) were found in Soumousso while Bama mosquitoes were all An. coluzzii. The first-generation female (F₁) of these mosquitoes was used for NIRS age analysis. Altogether, NIRS was unsuccessful determined mosquito chronological age in different environmental conditions (Fig. 6). The average difference between actual age and predicted age varied from 3.27 to 5.59 days. The lowest and highest average were found under end season conditions and insectary conditions, respectively.

Fig. 6.

NIRS accuracy to determine Anopheles chronological age under insectary, peak of the rainy season (August) and end of the rainy season/onset of the dry season (December) conditions programmed in incubators

Mosquito ability to transmit Plasmodium requires a minimum longevity covering the parasite extrinsic incubation period (EIP) estimated at 9 days [30]. Thus, the accuracy of the NIRS technique in classifying mosquito age as young (< 9 days) or old (≥ 9 days) under different environmental conditions was evaluated. Regardless of mosquito group and environmental conditions, the binomial predictive model indicated that NIRS was able to classify mosquitoes as young (< 9 days) and old (≥ 9 days), with 81% of accuracy for both ages’ classifications. The similar accuracies were found under each environmental condition with slight variations depending on the Anopheles group, ranging from 79 to 84% (Fig. 7). Additionally, although there is no significant difference, NIRS classified younger Anopheles (< 9 days old) better than older ones (≥ 9 days) whatever environmental conditions and mosquito group (Table 1). These accuracies were obtained by training the model on samples of one condition and validating on a subsample of the same condition. However, NIRS was unable to distinguish between young and old mosquitoes when the model derived from mosquitoes reared in one environmental condition was used to predict the age of mosquitoes reared in another condition (< 50% accuracy).

Fig. 7.

NIRS accuracy to classify Anopheles (An. coluzzii and An. gambiae s.l.) in two age classes (< 9 days or ≥ 9 days) under insectary, peak of the rainy season (August) and end of the rainy season/onset of the dry season (December) conditions programmed in incubators

Table 1.

NIRS accuracy to classify Anopheles age under different environmental conditions inside climate chambers

Within-sample accuracy (%)
Periods		An. gambiae s.l.				An. coluzzii
	Number	Accuracy	< 9	≥ 9	p	Accuracy	< 9	≥ 9	p
Insectary	342	84	87	80	0.35	84	84	83	1
Peak	288	83	86	79	0.32	82	82	81	0.83
End	333	79	81	77	0.83	81	83	79	0.79

Discussion

NIRS can successfully discriminate between laboratory-reared and field-collected An. gambiae s.s. and An. coluzzii with varying degrees of accuracy depending on the environmental conditions. Overall, this work revealed that NIRS was able to distinguish both An. gambiae s.s. and An. coluzzii in the laboratory and in the field with greater than 80% of accuracy. Inter-environmental conditions accuracies of species identification vary slightly in laboratory mosquitoes but more in field specimens. Nevertheless, wild mosquitoes caught in one season can broadly predict species in mosquitoes caught months later, though with some loss of accuracy. The considerable variation in inter-period accuracies could be due to several potential factors: the diversity of larval habitats, nutriments (blood and sugar sources), gonotrophic/physiological status [12] and resistance status to the insecticides of the mosquitoes [29]. Since these factors weren’t controlled in the field, the differences in accuracy couldn’t be solely attributed to the abiotic factors of various collection periods. Additionally, the spectra from laboratory mosquitoes differ from those of field mosquitoes in this study as reported by other authors [7]. This study underlines once more the risk in extrapolating laboratory findings to the real-life environment, as previously suggested by other authors [31]. Results also revealed that the model derived from peak season samples had the highest accuracy for predicting samples from any other conditions, both in the laboratory and in the field. This period is characterized by a low fluctuation in temperature and RH, which could have less impact on absorbances through the mosquito cuticle compared to other periods. However, since each model was trained on a relatively small number of samples with varying sizes, these results should be validated using larger number of mosquitoes. This is particularly important given that previous studies have shown that increasing sample size can enhance the NIRS accuracy [10]. In addition, the preprocessing techniques and advance machine learning such as artificial neural networks and autoencoders could also improve the NIRS accuracy [32]. Furthermore, the study showed that the machine learning algorithm trained using laboratory-mosquitoes reared under different environmental conditions predicted field-derived mosquito species with a low accuracy. The mosquitoes were reared in laboratory under controlled conditions, free from factors like competition, nutrition, parasitism and several other factors that could create differences between laboratory specimens and those in the wild. This may mean that laboratory-reared mosquitoes do not reflect all the variations found in wild populations, so that predictive models derived from them are able to correctly predict the species of field specimens.

The results revealed that NIRS was able to classify both An. coluzzii and An. gambiae s.l. of age < 9 days and ≥ 9 days with reasonable accuracies in each environmental condition. Although this technique is not useful to determine mosquito chronological age, it may provide a rapid means of determining an epidemiologically relevant measure of the proportion of the vector population that is enough old to transmit malaria. This is ecologically important because it may help in designing and monitoring vector prevention efforts and assess the risk of pathogen transmission in the human population. Overall, the analyses showed that the technique was slightly more accurate in predicting mosquitoes < 9 days old compared to their counterparts ≥ 9 days old. This supports the assertion that major biochemical changes detected by NIRS occur in mosquito cuticles from birth and stagnate between days 10 and 14 due to cuticle thickening [33]. Findings indicated that a calibration model derived from mosquitoes of specific environmental condition determined mosquito species of another condition with relatively good accuracy. However, classifying Anopheles age into two groups using a model trained on specimens from another environmental condition was challenging and resulted in low accuracy. This may have been driven in part by the low number of mosquitoes to train the models from some conditions or perhaps the diversity in spectra seen when temperature fluctuate. While statistical analyses require sufficient sample size to improve model accuracy [10], the number of mosquitoes using in this study substantially varied in the different seasons. Therefore, care should be taken interpreting the reliability of different models trained on data from different season. NIRS would be more useful if a calibration model derived from specimens in one environmental condition is able to predict age of Anopheles in another condition. This work suggests temperature and relative humidity could influence mosquito spectra and that future models should include some variability in these factors in the data used to train and validate the models. Similar results were reported by other authors who found that models derived from laboratory mosquitoes were unable to predict the age of field derived mosquitoes reared under ambient condition [7]. Similarly, models derived from one preserved mosquito were unsuitable for predicting the age of specimens stored in another preservative [34]. Based on these findings, it is suggested to estimate the age of mosquitoes from a given environmental condition using predictive models developed from mosquitoes reared under the same conditions. Ultimately, NIRS accuracy for Anopheles species identification and out-of-sample prediction should be improved by substantially increasing the number of mosquitoes used in the training and validating datasets. These mosquitoes should be reared in as diverse a set of environmental variables as possible to include the range likely in the wild mosquitoes being tested.

Conclusion

The findings revealed that NIRS is able to predict An. gambiae s.l. species under different environmental conditions with reasonable accuracy. In addition, models derived from An. gambiae s.l. in one environmental condition is able to predict Anopheles species in more other condition with relatively good accuracy. However, models trained using laboratory-mosquitoes reared under various environmental conditions predicted field-derived mosquito species with a low accuracy. NIRS provided a good classification of age under all environmental conditions for both An. coluzzii and An. gambiae s.l. when the age was classified into two groups age (< 9 days or ≥ 9 days) using binary classification models. The models derived from Anopheles in one environmental condition failed to predict the age of mosquitoes in other conditions. Further studies are needed to investigate the accuracy of NIRS for determining Anopheles infection status or Anopheles resistance to insecticide status under different environmental conditions in order to develop appropriate calibration of the machine before field application.

Abbreviations

NIRS

Near infrared spectroscopy

ITNs

Insecticide-treated nets

IRS

Indoor residual spray

RH

Relative humidity

PCR

Polymerase chain reaction

PLS

Partial least squares regression

ROC

Receiver operating characteristic curve

AUC

Area under curve

Author contributions

NDCD, DFD, and RKD conceived and designed the study. NDCD, DFD and BMS conducted the laboratory and field work. NDCD, LIGP and FC conducted the molecular analysis. NDCD, BMS and TSC were responsible for the data analysis. NDCD and DFD wrote the first draft of the manuscript. BMS, FC, JK, TSC and RKD reviewed the manuscript. All authors read and approved the final version of the manuscript.

Funding

The work was supported by the Centre d’Excellence Africain en Innovations Biotechnologiques pour l’Elimination des Maladies à transmission vectorielle (CEA-ITECH/MTV) of Nazi BONI University, Burkina Faso and UK Medical Research Council (MRC) Project Grant (MR/P01111X/1).

Availability of data and materials

The datasets are fully available without restriction from the corresponding authors by request. Further inquiries can be directed to the corresponding authors.

Declarations

Ethics approval and consent to participate

The project received approval from the local institutional ethics committee, reference number: A018-2017/CEIRES.

Consent for publication

“Not applicable” for that section.

Competing interests

The authors declare no competing interests.