Background
Effective entomological surveillance requires systematic collection, analysis, and interpretation of data on insects that transmit pathogens in different localities. It is essential for assessing risks and guiding the planning and implementation of vector control strategies, as well for monitoring, and evaluation of those strategies [
1]. The likelihood of pathogen transmission can vary widely, depending on factors such as the presence of competent vectors, favourable climatic conditions, the presence of vulnerable human populations and the presence of other vertebrate hosts, which may sustain the vector populations [
1]. Other factors may include the diversity of vector species in the area, their population dynamics, their behaviours in and around human dwellings such as the timing and location of biting, their resting behaviours and host preferences of these vectors [
2,
3].
Anopheles mosquitoes are considered particularly hazardous due to their propensity to feed on, and thus transmit pathogens to, humans, notably malaria, which causes approximately 620,000 deaths and about 250 million cases annually [
4]. Compared to mosquitoes from other regions, the Afro-tropical malaria vectors are particularly dangerous in this regard due to their comparatively greater preference for humans over other vertebrates [
2]. This attribute, which is generally estimated as the human blood index, has been considered an important measure of the stability of malaria in different settings [
5]; and is known to be highest in major malaria vectors, including
Anopheles gambiae, Anopheles funestus and
Anopheles coluzzii, which appear to be particularly well adapted synanthropes [
6]. Following closely is
Anopheles arabiensis, which can be an opportunistic vector species capable of blood-feeding readily on either humans or cattle, depending on availability [
2,
3,
7]. Consequently, while this behaviour poses a notable risk for the transmission of zoonotic pathogens in addition to malaria,
An. arabiensis is also a far less competent vector of malaria than either
An. gambiae,
An. funestus or
An. coluzzii [
8‐
11].
While anthropophagy (i.e. preference for feeding on humans) in malaria vectors can be augmented by the degree of endophily (i.e. preference for indoor resting), this behaviour can also be attenuated under high degrees of exophily (i.e. preference for outdoor resting). For example,
An. funestus is known for being both highly anthropophilic and highly endophilic [
2,
9], enforcing its major role in malaria transmission [
9,
10] though there are settings where it is known to bite outdoors early in the morning [
12,
13] or to feed on non-human hosts [
14]. On the other hand, mosquitoes that rest indoors are more likely to feed on human host, while mosquitoes that prefer to rest outdoors are more likely to feed on non-human host [
2,
15]. This might be due to mosquitoes feeding on the first host they encounter when presented with multiple hosts in the same environment [
7], or to the use of bed nets preventing access to human hosts [
16,
17]. Overall, accurate determination of the blood-feeding histories of malaria vector species is an important indicator of their feeding behaviour, their role in ongoing malaria transmission and the overall risk exposure of people within those settings.
Methods for investigating the blood-meal sources in mosquitoes include several techniques: the precipitin test observes the formation of a white precipitate resulting from the interaction between a saline extract of the blood meal and a suitable antiserum from a known host, indicating the presence of an antigen–antibody interaction [
18]; microsphere assays is a molecular-based assay involving uniquely labelled microspheres with host species-specific capture probes to detect host blood meals [
19]; microsatellite assays analyse short tandem repeat sequences in the mosquito’s DNA to identify blood sources based on unique genetic markers [
20]; enzyme-linked immunosorbent assays (ELISA) detect immunoglobulin G (IgG) from blood-fed mosquito samples [
21]; and polymerase chain reactions (PCR) target mitochondrial cytochrome b to identify arthropod blood meal sources [
22]. ELISA and PCR, the most common techniques for studying host blood meals in mosquitoes, have played a crucial role in understanding mosquito host preference since the early 1980s and emerged as powerful tools due to their sensitivity [
21‐
26]. These methods have evolved over time with modification to enhance accuracy and efficiency. ELISA, for instance, utilizes two basic procedures: indirect ELISA, where an antiserum is used to trap a particular IgG [
23], and direct ELISA, which relies solely on the antibody-enzyme conjugate to attach to host-specific IgG in the bloodmeal [
21,
24], currently preferred for its simplicity over indirect ELISA. PCR, being more sensitive due to specific primers targeting host DNA, has evolved from conventional PCR, which amplified human host DNA extracts at the human tyrosine hydrolase (TC-11 or HUMTHO1) and VWA (HUMVWFA31) [
25,
26], to the current multiplexed PCR capable of detecting five mammalian blood meals in mosquitoes in a single step (i.e., by size-differentiated DNA fragment on agarose gels) [
22]. While these techniques offer significant advantages, they also come with challenges such as being time-consuming, laborious, and require repeated use of expensive reagents, not always readily available in rural laboratories where field collections are conducted. Moreover, ELISA assays, one of the most widely used technique, are prone to high levels of cross-reactivity, occasionally failing to sufficiently distinguish between human and hon-human blood meals [
27]. Since field collections do not always yield synchronous physiological states, some of the blood meals may have been partly digested, which might also compound the detection capability of current methods [
28].
In a recent study, our team demonstrated that machine learning models trained on mid-infrared spectra data collected from mosquitoes fed on different hosts (4000 cm
−1 to 400 cm
−1 frequencies) (MIRS-ML) could accurately distinguish vertebrate blood meals in laboratory-reared
An. arabiensis mosquitoes without the need for molecular techniques [
29]. However, it was also noted that field validation would be necessary for multiple reasons. Firstly, in field settings, the time post-feeding is unknown, and the mosquitoes may have multiple blood meals, occasionally from multiple sources. Secondly, unlike laboratory settings where the age of mosquitoes is known, field mosquitoes vary in age and may have taken their 2nd, 3rd, or 4th meals. Thirdly, the amount of blood in the mosquito gut may be small in the field due to increased disturbance during feeding compared to controlled laboratory conditions, and lastly, the genetic variability for blood sources is higher in the field. Overcoming these challenges would enable the potential use of MIRS-ML in real-world field scenarios. We, therefore, concluded from the initial laboratory study that whereas the technique offers a unique opportunity to rapidly test individual mosquitoes for blood-type and other attributes, assessing blood-feeding histories of wild malaria mosquitoes would provide an opportunity to test its potential field validation.
The current study aimed to analyse the blood-feeding preferences of wild-caught malaria mosquitoes, by using MIRS-ML models to identify the sources of their blood meals. The study also explored how well the models trained using laboratory-reared mosquitoes can be applied to field-collected samples by incorporating specific transfer learning techniques previously used for predicting the species identification and age of mosquitoes collected in different countries [
30,
31]. The ultimate goal of the work was to demonstrate the utility of this approach for field applications. Implementing these models in the field would significantly enhance the knowledge of mosquito feeding behaviours and disease transmission, potentially informing more effective vector control strategies against multiple mosquito-borne diseases [
32‐
39].
Discussion
Human blood index (HBI), which reflects the tendency of mosquitoes to feed on humans compared to other vertebrates, is vital for assessing malaria transmission dynamics and the level of stability of transmission [
5]. Current techniques for determining mosquito blood meal sources are slow, labour-intensive, and expensive due to the need for costly reagents. They are also susceptible to errors, such as false positives from cross-reactivity with other antigens or due to the partial digestion of blood meals in the mosquito digestive system. Yet, as malaria endemic countries move towards elimination, there is a pressing need for simpler, more cost-effective methods that can be deployed at scale in malaria-endemic countries to improve entomological surveillance and evaluate the effectiveness of malaria control interventions.
This study demonstrates the first-ever field application of the simple mid-infrared spectroscopy and machine learning (MIRS-ML) approach for predicting the blood-feeding histories of malaria vector in rural Africa. Beyond this, the study also demonstrates the transferability of the laboratory-trained MIRS-ML models to identify and classify host blood meals in field-collected samples through the utilization of transfer learning techniques. For validation, PCR as the ‘ground truth’ was used to determine the actual blood-feeding histories of the field-collected mosquitoes; and examined a total of 1854 blood-fed Anopheles mosquitoes.
Based on the PCR analysis, most of the mosquitoes blood-fed on humans or bovines, and only a very small percentage had fed on other hosts, such as dogs and pigs. Given the inherent limitations of the PCR, classification of blood meals in 41% of the samples was impossible, possibly because they fed on a host other than those tested in this study and therefore could not be amplified with the primers used. Nonetheless, only mosquitoes confirmed to have fed on either humans or bovines were included in this analysis, as they were the vast majority; thus binary machine learning classifiers were trained for blood-meal prediction. The capability of the MIRS-ML models to classify mosquito blood-meal sources was demonstrated, achieving an accuracy of 88%, when using 338 spectra data collected from field samples (169 human-fed and 169 from bovine-fed mosquitoes). This demonstrates a realistic opportunity to deploy such simple methods for estimating HBI, thereby extending the capability of infrared-AI based systems already well demonstrated for tracking several other entomological attributes [
45].
In prior work using age-synchronized laboratory-reared mosquitoes, the focus was on predicting blood-meal sources with
An. arabiensis, where the MIRS-ML approach achieved a classification accuracy of–98% for four blood meal sources (bovine, human, goat and chicken) [
29]. Whereas the mosquitoes used in that earlier study were only 6–8 h post-feeding, this current study included a broader range of age groups and natural variation in the degrees of digestion of the bloodmeals. This current study therefore strongly demonstrates the potential of the MIRS-ML approach for realistic field surveillance, even when the time of actual blood-feeding and digestion stages is unknown upon sample collection and preparation.
A major achievement in the present work is the demonstration of the transferability of laboratory-trained models to field samples through the application of transfer learning. The transferability of laboratory-trained models achieved a classification accuracy of 90% in predicting blood-meal sources for field-collected
An. funestus. The base laboratory model was initially trained using spectra data from blood-fed
An. arabiensis [
29], which was then augmented by incorporating a small subset (
n = 100, with 50 samples each from humans and bovine blood-fed
An. funestus spectra) of field-collected data into the model. This implies that the technique can be extended to assess blood-meal sources in the abdomens of Afrotropical malaria vectors, as the species would not be a confounding factor in this case. It also implies that the generalizability of this model will cut across laboratory and field sample prediction, and therefore, sample origin might not be a confounding factor. Since field-collected mosquitoes were likely of varied ages, and therefore mosquito age, a factor readily classifiable by MIRS-ML models [
30], is also unlikely to be a confounder, and can be overcome by similar transfer learning approaches. The results presented here corroborate with previous studies in which the utilization of transfer learning successfully generalized predictions of mosquito age and species across different countries and laboratories [
30,
31]. This approach effectively accounts for the inherent variability of mosquitoes from different environmental and ecological settings or genetic backgrounds, which could otherwise limit the generalizability of ML models trained on mosquito spectra data to new mosquito populations. Indeed, the genetic variability for blood meals in the field is likely high, and blood-fed mosquitoes collected during the study contained a mixture of fully engorged and partially consumed blood meals.
Partial digestion or low quantity of ingested blood meals, could potentially impair the capability of MIRS-ML to accurately identify or differentiate between various blood meals, thereby affecting the Human Blood Index (HBI) estimates. To mitigate this, it is advised against including gravid mosquitoes in samples and recommended to preserve all blood-fed mosquitoes immediately upon collection to halt any biochemical changes before spectroscopy. Currently investigating this phenomenon, preliminary studies have demonstrate a notable decrease in MIRS-ML accuracy after 36 h post-feeding (Mgaya et al. (unpublished), which coincides with gravidity in a typical 2–3 day gestation period under optimal conditions. In this paper, field models closely aligned with PCR outcomes, considered as the benchmark, despite the inability to precisely determine the gestational stage of mosquitoes at the time of collection each morning post-trapping. Moreover, earlier studies by Mukabana et al. [
28], have successfully used PCR to amplify host DNA up to 32 h post-feeding after which the host DNA is degraded. Crucially, the analysis only incorporated samples that yielded successful PCR amplification of host DNA for MIRS-ML training, discarding all non-amplified samples. This selection criterion may inadvertently introduce bias since the partially or fully digested blood meals may be the ones least likely to yield good-quality host DNA. Future models should therefore include samples of mosquitoes that have blood-fed on known hosts, 1–4 days post-feeding to evaluate the efficacy of MIRS-ML across various stages, including gravid and post-oviposition states. Lastly, though the model was already trained on a large number of mosquitoes, it is recommended to increase these sample sizes and obtain mosquitoes from different sampling locations so as to neutralize effects such as partial blood-meals and partial digestion, as well as any effects of environmental or microclimatic factors affecting blood feeding and digestion.
Indeed, increasing the number of field samples for transfer learning not only enhanced the classification accuracy for field blood-fed mosquitoes but also improved the precision in estimating the HBI in comparison to the ‘ground truth’ PCR method. This indicates that the technique has the potential to be a reliable method for estimating HBI, capable of generalizing HBI estimations in field-collected mosquitoes as effective as PCR. Therefore, it can provide valuable information to national malaria control programs regarding the feeding preferences of malaria mosquitoes.
Despite the successes of this technique, there remain several gaps. Firstly, it is unclear whether the technique can detect mixed blood meals, a situation that is more likely to occur in the field, remains unanswered, warranting future investigation. Secondly, PCR and ELISA remains highly sensitive and specific, known for their accuracy in detecting host DNA and specific protein from blood meals, even in small amounts, respectively. Although MIRS-ML has demonstrated notable accuracies in detecting mosquito blood meals, its performance, being highly sensitive and specific, depends on the quantity and quality of the training data and machine learning algorithms used. This robustness of the model will contribute to its ability to handle variations. Thirdly, the machine learning models in this study were trained using An. funestus mosquitoes that had blood-fed on humans and bovines. This choice was made because most mosquito samples collected from the field contained either human or bovine blood in their abdomens, while only a minority had dog blood or mixed blood-meals. Consequently, the available samples were insufficient to adequately train the machine learning models to detect mosquito blood-meal sources from hosts other than humans and bovines. In their current state, these models would face challenges in field deployment since they will not be capable of identifying blood-meal sources from other potential hosts often found in human dwellings such as goat, pig, and chicken. However, considering that the transferability of the laboratory-trained models for field sample prediction has also been demonstrated, the deployment of these models could involve initially training them on laboratory data, which can be generated in large quantities. Additionally, this approach allows for the inclusion of a wider range of hosts, ensuring accurate mosquito blood-meal source prediction from all common hosts typically found near human dwellings, including humans, bovines, goats, dogs, pigs and chickens. Thus, once validated, MIRS-ML approaches have the potential to make significant contributions to understanding the dynamics of disease transmission involving humans, livestock, wildlife, and vectors. Specifically, they could offer valuable insights into scenarios where mosquitoes have opportunities to feed on multiple host species.
Interestingly, despite its anthropophilic behaviour,
An. funestus, the main vector in the study area, was found to also blood-feed on bovines. This finding is consistent with previous studies that demonstrated a potential switch in host choice by
An. funestus from humans to cattle [
46,
47]. In brief, given the circumstances of the collections, this observation may be explained by several factors: Firstly, the houses where mosquito collections were conducted had been supplied with intact bed nets before the collections started, which might have created a physical barrier, reducing mosquito exposure to humans [
48]; and forcing mosquitoes to use alternative blood sources in the surrounding areas as previously documented by Iwashita et al. [
48]. Secondly, it might have been a result of the zoopotentiation effect, which refers to the increased tendency of mosquitoes to feed on humans living near livestock [
49,
50], especially when livestock in close proximity to human dwellings emit heat and odour cues that attract mosquitoes. In such circumstances, not only do zoophagic mosquitoes find additional blood sources that they already prefer, but even the naturally anthropophagic mosquitoes may also accidentally feed on cattle when host cues become mixed nearby. There is a lot of evidence suggesting that zoopotentiation may increase malaria transmission risk by creating an alternative source of bloodmeals, consequently increasing both mosquito survival rates and abundance [
51‐
55]. This interaction of mosquitoes between humans and non-human hosts may also elevate the likelihood of transmitting parasitic helminths and zoonotic pathogens [
32‐
39,
56].
Infrared spectroscopy and machine learning methods have already been demonstrated for several other use cases, such as age-grading mosquitoes [
30,
31,
57‐
59], detection of pathogens inside mosquitoes [
60], identification of mosquito species [
30] and even detection of parasites in human blood [
61‐
63]. This demonstration of its usefulness for analysing the blood-feeding histories of mosquitoes in both the laboratory (as previously shown [
29]) and the field (this current study), underscores the unique potential of the technology as a one-stop system for comprehensive analysis of entomological and parasitological indicators of malaria and other mosquito-borne diseases.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.