Infrared-based techniques are increasingly being tried for entomological studies such as age-grading and species identification of malaria vectors [19, 27, 30–32]. As these techniques are still in early-stage development, researchers have mostly depended on specific sample handling approaches, with limited considerations for either standardization or alternatives. However, for NIR-specific uses, it has been demonstrated that the methods can work with mosquitoes preserved using multiple techniques. Dowell et al., for example, demonstrated that the chronological ages of mosquitoes could be predicted from NIR-spectra to within 1.4 days when using either desiccants, ethanol, Carnoy, RNA later®, or refrigeration [26]. Separately, Sikulu et al, showed that preserving mosquitoes in RNA later® reduced the likelihood of misclassifying the age of Anopheles mosquitoes, further emphasizing the potential of this preservative [25]. These studies demonstrated the expanded potential for infrared-based applications and provided a basis for additional investigations.
Since the infrared spectroscopy applications now increasingly also include the mid-infrared spectral range, studies on different sample handling techniques should be expanded to the MIR application as well. This current study therefore investigated the effects of different preservation methods on the performance of a previously implemented mid-infrared based approach for age-grading female malaria vectors[20, 28]. The study also tested whether different durations of sample storage durations could influence the performance of the age-grading models based on mid-infrared spectra. By using laboratory-reared mosquitoes that were either young (5-days old) and old (17-days old), this study also enabled the testing of these methods for age-classification, independent of preservation methods or storage durations.
Desiccation over silica gel is a common preservation method used by entomologists and has also been widely used in previous infrared-based studies, thus it was considered in this study as the primary reference, against which other methods were compared. Similarly, the storage durations of 1,4 and 8 weeks were selected to represent a practical range over which samples would normally be stored before analysis even in cases where there are limited analytical resources. One week was considered as the baseline, against which other durations could be evaluated.
Of the seven classification models tested, support vector machines (SVM) were the best performing at predicting the ages of mosquitoes preserved in different preservation methods, achieving about 96% accuracy. Broadly, the MIRS-based approach could accurately classify the age groups of mosquitoes even after 8 weeks of storage, even though performance was best for samples stored in 1 week. When data was put together for all storage durations, and the SVM model trained using just samples preserved in silica, the highest accuracy was obtained when the unseen data being predicted was also from silica-preserved mosquitoes. However, when this model was used to predict samples preserved in other methods, the accuracy declined significantly, suggesting the need to standardize the preservation method between the training and validation sets of samples. The same observation was made when the model was trained using samples preserved by silica for one week and then used to perform the predictions for other preservations and storage durations. Here too, the highest predictions accuracies were obtained from the preservation method and storage duration that was used as a reference (Table 2).
While these data do not necessarily offer a comprehensive analysis of all possible preservation methods, they clearly demonstrate the need to either standardize the preservation methods or at least deploy an additional layer of statistical procedures such as transfer learning [31], where a small amount of data from the test populations is introduced into the training set, to neutralize the differences introduced by using different preservatives. Such statistical approaches have been applied to improve predictions on samples collected in different countries or laboratories for mosquito age grading [20]. In the study conducted by Siria et al, the transfer-learning approach was demonstrably effective at extending the utility of the deep learning models to predict the ages of field-collected mosquitoes in both Tanzania and Burkina Faso [20]. Perhaps the most practical option would simply be to require standardized treatments of test and training samples in both laboratory and field studies.
Silica gel has been used by researchers in preserving many samples for both species identification, age grading and blood meal experiments at a relatively low cost [20, 22, 27, 32]. This makes it ideal for storing large numbers of mosquito samples in field settings over long durations. This current study has also shown that silica gel desiccation may be ideal for storing mosquito samples for a short period at ~ 5°C, as this were the samples for which age classification was most accurate. On the contrary, pre-frozen samples achieved far lower age predictions across the different storage durations, when the SVM model was trained on silica gel as a reference (Table 2). These samples were not as dry as those preserved in silica gel for the same duration; therefore, excess water content may have limited the full potential of the machine learning predictions of the spectral data even after the data was cleaned. Because of the moisture content even after the drying period before scanning, the frozen samples were easily crushed by the anvil and provided no resistance when pressed against the ATR crystal of the spectrometer. This may add additional complexity to data analysis, requiring that certain data points are discarded due to excessive water content as previously suggested by González et al., [28]. Lastly, ethanol has also been used widely for mosquito preservation, especially where the nucleic acid component is needed for further analyses [33]. In this study, the prediction accuracies of models trained with ethanol-preserved samples dropped to 50% when age-classifying silica-preserved mosquitoes and to 56% when age-classifying frozen samples. Further analysis, beyond the scope of this current study, may be needed to evaluate these comparisons, and possibly include additional statistical approaches.
This study also allowed direct assessment of whether variations in storage duration can impact the accuracy of mid-infrared-based approaches for mosquito age-grading. Here, the samples stored for one-week were initially used as the standard reference and used to train the basic machine learning models, which were then used to predict age classes of mosquitoes stored for different durations. It is particularly important to evaluate these differences since entomological surveys are typically time-consuming and can generate very large numbers of samples that cannot be analysed on the same day immediately. As a result, some form of extended storage for weeks or months is often necessary, especially where the equipment for sample analysis does not exist on-site.
Overall, these results suggest that Mid-infrared spectroscopy coupled with machine learning can predict the age of mosquito species stored in different preservation methods for different periods of up to 8 weeks. As previously demonstrated, the approach has the advantages of being quick to perform, cost-effective, and reagent-free, this being reliable even in low-resource settings [20, 28].
One limitation of this study was that it was carried out for only one species of malaria vectors, and therefore more research is needed to investigate if there will be any variations in other mosquito species transmitting malaria. Moreover, the range of preservatives and durations of storage were limited to just three each, to ensure feasibility. Further investigation may reveal that such models may respond differently when using an expanded range of preservatives or storage durations. Nonetheless, the data shows that silica-based preservation is a satisfactory starting point for samples destined for spectroscopy; and can be used for several weeks of storage.