Inferring insect feeding patterns from sugar profiles: a comparison of statistical methods

• Published in: Ecological entomology 2021-02, Vol.46 (1), p.19-32
• Main Authors: Luquet, Martin, Parisey, Nicolas, Hervé, Maxime, Desouhant, Emmanuel, Cortesero, Anne‐Marie, Peñalver‐Cruz, Ainara, Lavandero, Blas, Anton, Sylvia, Jaloux, Bruno
• Format: Article
• Language: English
• Subjects: Agricultural sciences; Algorithms; Applications; Carbohydrates; classification; Confidence; Datasets; Feeding; feeding inference; Food sources; Insects; Life Sciences; machine learning; Nutritional ecology; prevalence estimation; Sciences and technics of agriculture; Statistical methods; Statistics; Sugar
• ISSN: 0307-6946; 1365-2311
• DOI: 10.1111/een.12971

Investigations in nutritional ecology often require the identification of animal feeding patterns in natural conditions (what, where, and when do animals eat). Thus, methods are needed to trace not only individual resource uptake but also the relative use of different resources in a population or community. Recent biochemical developments allow predicting the use of sugar‐rich resources from insects in the field. Individual feeding status (feeding history, food sources) is inferred by comparing insect sugar profiles with those of individuals fed on controlled diets. Individual assignations are then used to predict the relative consumption of different resources at the population or community level. As both steps may generate error, accurate prediction rules are needed. However, research from other domains (e.g., protein‐marking studies) suggests that classical decision rules used for such tasks may sometimes induce bias. This study evaluated the performance of these rules and compared them to alternative methods on simulated, realistic datasets. It tested different methods for individual classification but also introduced methods for prevalence estimation, whose specific purpose is to estimate the relative frequency of different classes. Alternative methods substantially outperformed the traditional algorithms to predict insect individual feeding status and population class distribution (relative frequency of insects with different feeding status). This study provided a simple decision tool to choose a method according to dataset size, variance, and biochemical method used. Alternative methods should increase prediction confidence in future studies. Such approaches should easily be generalized to a wider range of systems. Supervised learning algorithms are effective tools to predict insect feeding from sugar profiles in the field. Random Forests are the most performant method to predict individual feeding status in many situations. Counting of individual assignations can be adjusted to better estimate feeding patterns at the population scale.