Machine learning prediction of multiple anthelmintic resistance and gastrointestinal nematode control in sheep flocks

Predição de resistência anti-helmíntica múltipla por aprendizado de máquina e controle da verminose em rebanhos ovinos

Abstract

The high prevalence of Haemonchus contortus and its anthelmintic resistance have affected sheep production worldwide. Machine learning approaches are able to investigate the complex relationships among the factors involved in resistance. Classification trees were built to predict multidrug resistance from 36 management practices in 27 sheep flocks. Resistance to five anthelmintics was assessed using a fecal egg count reduction test (FECRT), and 20 flocks with FECRT < 80% for four or five anthelmintics were considered resistant. The data were randomly split into training (75%) and test (25%) sets, resampled 1,000 times, and the classification trees were generated for the training data. Of the 1,000 trees, 24 (2.4%) showed 100% accuracy, sensitivity, and specificity in predicting a flock as resistant or susceptible for the test data. Forage species was a split common to all 24 trees, and the most frequent trees (12/24) were split by forage species, grazing pasture area, and fecal examination. The farming system, Suffolk sheep breed, and anthelmintic choice criteria were practices highlighted in the other trees. These management practices can be used to predict the anthelmintic resistance status and guide measures for gastrointestinal nematode control in sheep flocks.

Keywords:
CARTs; machine learning; multidrug resistance; gastrointestinal nematodes; random forest

Resumo

A criação de ovinos é afetada pela alta prevalência de Haemonchus contortus e pela resistência anti-helmíntica, cujas complexas inter-relações podem ser elucidadas por técnicas de aprendizado de máquina. Árvores de classificação identificaram a importância de práticas de manejo em 27 rebanhos ovinos, para a predição da resistência anti-helmíntica múltipla. A resistência a cinco anti-helmínticos foi determinada pelo teste de redução de contagem de ovos nas fezes (TRCOF), e as 20 propriedades com eficácia inferior a 80% no TRCOF para quatro ou cinco anti-helmínticos foram consideradas resistentes. Os dados foram aleatoriamente separados em 1.000 amostragens de treinamento (75%) e teste (25%), e as árvores de classificação foram geradas para os dados de treinamento. Dessas, 24 árvores (2,4%) apresentaram acurácia, sensibilidade e especificidade de 100% para a predição do estado de resistência nos dados de teste. O tipo de forragem foi detectado em todas as árvores, enquanto a área de pastagem e a realização de exame de fezes foram observadas nas árvores mais frequentes (12/24). O sistema de produção, a raça ovina Suffolk e o critério de escolha do anti-helmíntico geraram quebras nas demais árvores. Essas práticas de manejo podem predizer o estado de resistência anti-helmíntica em rebanhos e orientar medidas, visando ao controle da verminose em ovinos.

Palavras-chave:
CARTs; aprendizado de máquina; resistência múltipla; nematoides gastrintestinais; random forest

Introduction

The economic importance of sheep in the supply of animal-origin proteins mainly relies on the use of small areas unsuitable for cattle production and agriculture (Sargison, 2012). According to recent estimates, Brazilian sheep flocks, composed of 20.5 million heads (IBGE, 2022), cannot meet the national consumer demand, which depends on lamb importation (FAO, 2022). Therefore, it is necessary to support, develop, and invest in sheep production systems in Brazil to increase productivity and profitability.

In tropical countries, the main challenges in sheep farming are adverse climatic conditions, competition for resources, including water and food, and the high prevalence of gastrointestinal nematodes (McManus et al., 2011). Among them, Haemonchus contortus is the most prevalent and pathogenic parasite in sheep (Waller, 2004), leading to losses estimated at US$ 107.5 million per year (Chagas et al., 2022). These losses are secondary to reduced growth and weight gain, inferior meat and wool quality, high costs of therapeutics, and animal death, especially in the case of resistance to anthelmintics used for nematode control (Costa et al., 2007; Miller et al., 2012).

Despite of the large number of commercial products, only four classes of broad-spectrum anthelmintics, i.e., benzimidazoles, imidazothiazoles, salicylanilides, and macrocyclic lactones, are available for treating sheep in Brazil (Chagas et al., 2013). Free access, ease of use, and lack of technical guidance have led to massive and incorrect use of anthelmintics, reducing their efficacy and promoting multidrug resistance (Doyle & Cotton, 2019). Because anthelmintic resistance leads to treatment failure, it can be detected in vivo through a fecal egg count reduction test (FECRT) after anthelmintic treatment of sheep hosts (Coles, 2005).

The establishment of resistance is affected by management practices used in flocks (Barger, 1997). Thus, resistance can be delayed by reducing pasture contamination and sheep infection not only through the use of anthelmintics but also by controlling environmental conditions (Bath, 2011; Leathwick, 2012; McBean et al., 2016; McFarland et al., 2022).

Similarly to other health traits, resistance data are complex, unbalanced, and contain missing values, resulting in a nonlinear complex interrelationship between the response and explanatory variables (Speybroeck, 2012). Thus, statistical and machine learning tools are suitable for detecting the factors that lead to anthelmintic resistance. Among them, classification and regression trees (CARTs) are non-parametric techniques that explore the interactions among variables, remove irrelevant covariates, produce visual results that are easy to interpret, and can predict the response to new observations (Speybroeck, 2012; Izbicki & Santos, 2020). A classification tree is built with recursive partitioning of the explanatory variables (resulting in nodes), minimizing the errors (the lower the error, the larger the branches and the importance of the variable to predict the outcome), and explaining, as much as possible, the categorical response variable in the leaves (Speybroeck, 2012; Izbicki & Santos, 2020). The random forest approach aggregates a collection of random decision trees, aiming to optimize a predictor and explore all possible tree predictors simultaneously, usually resulting in better performance (Genuer & Poggi, 2019).

Recently, machine-learning techniques have been used in veterinary parasitology. They were used to predict the resistance of sheep against gastrointestinal nematodes (Freitas et al., 2023b), to classify Famacha score anemia based on automatic analysis of ocular conjunctiva images (Freitas et al., 2023a), to perform fecal Strongyle egg counts from video footage (Bucki et al., 2023), to detect environmental risk factors affecting haemonchosis incidence based on egg per gram counts (Suresh et al., 2022), and as a potential approach for anthelmintic drug discovery (reviewed by Zamanian & Chan, 2021). To the best of our knowledge, this is the first study to use machine learning to predict anthelmintic resistance in flocks based on management practices.

In a previous study using the same data (Niciura et al., 2012), risk factors in management practices were identified using the frequency of genetic resistance to benzimidazoles in H. contortus as the response variable. However, as most flocks were resistant to other anthelmintic classes besides benzimidazoles, using FECRT data as the response variable and applying machine learning approaches have the potential to identify the factors involved and their interrelationships with the establishment of multiple anthelmintic resistance.

Thus, knowledge of parasite epidemiology, resistance status, and management practices in flocks can be used to develop new strategies for parasite control. In this study, classification trees were built based on 36 management practices and FECRT efficacy data for five anthelmintics (albendazole, closantel, ivermectin, levamisole, and moxidectin) in 27 sheep flocks in São Paulo State, Brazil. The objective of this study was to generate classification trees and identify relevant management practices that lead to resistance to multiple anthelmintics.

Material and Methods

A survey was conducted among 34 sheep farmers in São Paulo State, Brazil, from 2008 to 2010 to obtain information on 36 management practices related to infrastructure, health, and feed management (Veríssimo et al., 2012). In addition, an FECRT for a control group and five anthelmintics (albendazole, closantel, ivermectin, levamisole, and moxidectin) assessed the reduction in fecal egg counts (FEC) (Ueno & Gonçalves, 1998) 10-14 days after treatment, obtaining anthelmintic efficacy in 27 flocks (Veríssimo et al., 2012). The following formula was used for calculation of anthelmintic efficacy in FECRT:

$\%Efficac{y}_{AH}=\frac{meanFE{C}_{Control}+meanFE{C}_{AH}}{meanFE{C}_{Control}}x100$ (1)

where Efficacy_AH is the efficacy of each anthelmintic; meanFEC_Control is the mean fecal egg count for the non-treated control group 10-14 days after FECRT; and meanFEC_AH is the mean fecal egg count for each anthelmintic-treated group 10-14 days after FECRT. An anthelmintic efficacy lower than 80% in FECRT indicate high resistance (Voigt et al., 2022).

In total, 1,000 classification trees were built from the 27 sheep flock dataset to predict the development of in vivo multidrug resistance (as the categorical response variable) based on management practices (explanatory variables) used in each flock. Considering the quartile distribution of anthelmintics with high resistance, flocks were classified as resistant (1) when presenting 4 and 5 anthelmintics with FECRT < 80% (20 flocks), and as susceptible (0) when presenting 0 to 3 anthelmintics with FECRT < 80% (7 flocks).

All quantitative (total area, grazing pasture area, time in farming, number of dams, number of sires, and number of heads) and categorical (wetland, flooring, main income source, secondary exploration, grazing of cattle, farm accounting, technical assistance, frequent animal incorporation, quarantine, Dorper, Ile de France, Santa Inês, Texel and other breeds, crossbred, region of animal origin, forage species, rotational grazing, farming system, shared grazing, anthelmintic choice, anthelmintic rotation, combination of drugs, deworming schedule, dose-and-move practice, estimation of weight, FEC examination, Famacha, FECRT, and resistance status) variables used in the analysis and their descriptive statistics in the 27 flocks are presented in Tables 1 and 2.

Table 1
Descriptive statistics for quantitative variables of management practices and anthelmintic resistance in 27 sheep flocks from São Paulo State, Brazil.

Variable	Min	1st Quart	Median	Mean	3rd Quart	Max	NA
Total area (ha)	7.5	33.0	100.0	273.4	217.0	2000.0	2
Grazing pasture area (ha)	2.0	9.5	27.2	49.0	63.2	194.0	3
Time in farming (years)	2.0	4.3	5.0	8.6	8.0	51.0	1
Number of dams	50.0	133.2	241.0	397.8	450.0	1500.0	3
Number of sires	1.0	4.0	12.5	46.1	22.5	773.0	3
Number of heads	95.0	183.0	316.0	594.1	742.0	2028.0	2
Anthelmintics with FECRT < 80%	1.0	3.5	4.0	3.9	5.0	5.0	0

• FECRT = fecal egg count reduction test; Quart = quartile; Min = minimum; Max = maximum; NA = missing values.

Table 2
Descriptive statistics for categorical variables of management practices and anthelmintic resistance in 27 sheep flocks from São Paulo State, Brazil.

Variable	Category 1	Category 2	Category 3	Category 4	Category 5	Category 6	Category 7
Wetlands	17 (no)	10 (yes)
Flooring	24 (concrete)	2 (slat)	1 (NA)
Main income source	19 (no)	8 (yes)
Secondary exploration	16 (no)	9 (reproduction)	1 (wool)	1 (milk)
Grazing of cattle	17 (yes)	10 (no)
Farm accounting	16 (yes)	11 (no)
Technical assistance	13 (frequent)	6 (sporadic)	8 (no)
Frequent animal incorporation	21 (no)	6 (yes)
Quarantine	21 (yes)	5 (no)	1 (NA)
Dorper breed	19 (no)	7 (yes)	1 (NA)
Ile de France breed	21 (no)	6 (yes)
Santa Inês breed	22 (yes)	5 (no)
Suffolk breed	17 (no)	10 (yes)
Texel breed	20 (no)	7 (yes)
Other breeds	22 (no)	5 (yes)
Crossbred	14 (yes)	12 (no)	1 (NA)
Region of animal origin	13 (SE)	6 (NE)	4 (SE + NE)	1 (SE + CW)	1 (SE + S)	1 (NE + S)	1 (NA)
Forage species	7 (Brachiaria)	6 (Cynodon)	5 (Panicum)	3 (Brachiaria + Panicum)	3 (Brachiaria + Panicum + Cynodon)	3 (Brachiaria + Panicum + Cynodon + Pulse)	2 (other)
Rotational grazing	22 (yes)	5 (no)
Farming system	22 (semi-intensive)	3 (intensive)	2 (extensive)
Shared grazing	14 (no)	7 (cattle)	5 (horses + cattle)	1 (horses)
Anthelmintic choice	15 (technician)	4 (seller)	3 (efficacy)	2 (other)	1 (technician + efficacy)	1 (experience)	1 (NA)
Anthelmintic rotation	11 (based on efficacy)	8 (after each treatment)	5 (based on FECRT)	3 (NA)
Combination of drugs	17 (no)	10 (yes)
Deworming schedule	8 (target + strategic)	7 (target)	6 (fixed)	3 (fixed + target + strategic)	1 (strategic)	1 (fixed + strategic)	1 (fixed + target)
Dose-and-move practice	16 (no)	11 (yes)
Estimation of weight	16 (visual)	8 (scale)	3 (scale + visual)
FEC examination	11 (sporadic)	11 (no)	3 (monthly)	1 (annual)	1 (quarterly)
Famacha	18 (yes)	9 (no)
FECRT	22 (no)	5 (yes)
Resistance status	20 (resistant)	7 (susceptible)

• FEC = fecal egg counts; FECRT = fecal egg count reduction test; NA = missing information; SE = Southeast; NE = Northeast; CW = Center-West; S = South.

The data were split into 1,000 random training (75%) and test (25%) sets, and trees were built using raw data (Steinberg, 2009) with the recursive partitioning (rpart) package (Therneau & Atkinson, 2022) in R version 4.2.1 (R Core Team, 2022). The parameters included method = “class”, split = “gini”, cp = 0, minsplit = 1, maxdepth = 30, and xval = 10, and the trees were plot with treemisc (Greenwell, 2022). The accuracy of prediction was calculated from the confusion matrix, and the package caret (Kuhn, 2022) was used to calculate the area under the receiver operating characteristic (ROC) curve (AUC), sensitivity, and specificity.

In addition, a random forest approach was used to confirm the results obtained from the classification trees. To this aim, missing data were replaced with the mean (for quantitative variables) or mode (for categorical variables). Then, a random forest (Liaw & Wiener, 2002) model, including the parameters importance = T, proximity = T, ntrees = 1000, and sample size was fitted to the training data (75%) and validated using the test data (25%).

Results

From the 1,000 classification trees fitted to the training data, 24 trees predicted the anthelmintic resistance status in the test group with 100% accuracy, sensitivity, and specificity and an AUC of 1, and the ROC curves for training and test data are shown in Figure 1. Forage species, grazing pasture area, and FEC were the explanatory variables predicting multidrug resistance in the most frequent (12/24) trees (Figure 2A). According to these trees, farms that have an exclusive pasture of Cynodon or Panicum forage species (right branch) and a grazing pasture area ≥ 9.6 ha are susceptible, whereas, when other forage species (left branch) are present, they are susceptible only if they perform FEC examination every month (Figure 2A). In the second most frequent trees (8/12) (Figure 2B), the farming system was included in addition to forage species, FEC exams, and pasture grazing areas. Farms with pasture areas < 10 ha are susceptible only if they adopt intensive farming systems. Other management practices were detected in the less frequent trees, such as anthelmintic choice criteria based on both technical recommendations and efficacy, resulting in susceptibility (Figures 2C and 2D), and raising the Suffolk sheep breed, resulting in anthelmintic resistance (Figure 2E).

Figure 1
Area under the receiver operating characteristic (ROC) curve for prediction of multiple anthelmintic resistance in 27 sheep flocks with classification trees using training (75%) and test (25%) data.

Figure 2
Classification trees of multiple anthelmintic resistance based on forage species, fecal egg count exam (FEC), grazing pasture area in ha, anthelmintic choice criteria, and Suffolk sheep breed in flocks of São Paulo State, Brazil. Cyn: Cynodon, Pan: Panicum, 1m = monthly, Tec+Effic = decision for anthelmintic treatment based on technical recommendation and efficacy.

Among the 24 classification trees, 24 out of the 36 investigated management practices were identified as important for the prediction of anthelmintic resistance. The importance scores were then retrieved, and the variables were ordered based on the sum of their scores (Table 3), revealing the grazing pasture area, forage species, number of heads, number of dams, and FEC examination as the top five variables for the prediction of multidrug resistance.

Table 3
Sum of scores for important management practices for classification trees built to predict anthelmintic resistance in sheep flocks.

Variable	Importance score sum
Pasture area	80.56
Forage species	69.23
Number of heads	50.66
Number of dams	40.45
FEC exam	39.09
Total area	35.93
Region of animal origin	29.08
Deworming schedule	28.75
Time in farming	21.00
Farming system	16.87
Combination of anthelmintics	12.83
Shared grazing with other species	12.69
Famacha exam	10.51
Dose-and-move practice	9.82
Grazing of cattle	8.86
Anthelmintic choice	8.61
Number of sires	8.19
Santa Ines breed	6.57
Texel breed	5.26
Frequent animal incorporation	3.45
Main income source	3.00
Suffolk breed	2.40
FECRT	1.06
Technical assistance	0.60

The random forest model resulted in a prediction with 86% accuracy, 100% sensitivity, and 50% specificity. Despite the inferior results compared to the classification trees, the ten most important variables (Figure 3), based on node impurities, are among those highlighted in the trees.

Figure 3
Importance of management practice variables based on the random forest model to predict anthelmintic resistance in sheep flocks.

Discussion

In well-managed flocks, fewer unnecessary treatments are applied to animals; consequently, selective pressure on the development of anthelmintic resistance is reduced. In this study, several management practices applied to sheep flocks were able to predict anthelmintic resistance status. The most important practices highlighted in the classification trees were forage species, grazing pasture area, FEC examination, farming system, criteria for anthelmintic choice, and Suffolk sheep breed.

Higher anthelmintic susceptibility was detected in farms with exclusive pastures of Cynodon and Panicium, whereas the combination of these species with other forage and pulse species resulted in resistance. Thus, handling only one type of forage may be easier than handling multiple different exigencies. Additionally, the forage itself may influence anthelmintic resistance in several ways. First, short-cropped grass accumulates more concentrated larvae than taller grass (Bath, 2011). Second, forage nutritional quality influences animal nutrition and physical condition, which, in turn, affects the effectiveness of the immune response and mechanisms to support the parasite load (reviewed by Walkden-Brown & Kahn, 2002; reviewed by McRae et al., 2015). Brachiaria humidicola and Panicum maximum cv. Massai had low alimentary values, whereas B. brizantha cv. Marandu had a high forage quality and Cynodon cv. Coastcross and P. maximum cv. Colonião, Tanzânia, and Tobiatã presented high nutritional and alimentary values (Corrêa & Santos, 2009). Nutritional variations among forage species may have influenced the response of sheep to nematodes and infection levels. Additionally, the morphology and structure of plants interfere with nematode larval migration from the feces in the soil to the forage tip ingested by animals and larval survival as a consequence of solar radiation exposure and desiccation. Seasonal temperature and humidity conditions affect H. contortus larval migration in B. decumbens (Santos et al., 2012), and a large number of infective larvae were recovered from U. humidicola and M. maximus cv. Massai compared with U. brizantha cv. Marandu and M. maximus cv. Mombaça (Pires et al., 2021). However, no differences in larval migration were observed among the Stylosanthes spp., B. brizantha cv. Marandu, B. brizantha cv. Xaraes, and P. maximum cv. Tanzânia (Oliveira et al., 2009; Urzedo et al., 2022). Thus, the availability of larvae in pastures being ingested by the animals and affecting infection rates may be influenced not only by the species (which was the information retrieved in the surveys) but also by forage cultivars.

In sheep grazing pasture areas, anthelmintic susceptibility associated with larger areas can be related to the increase in food supply and consequent improved animal nutrition, dilution of nematodes in the pasture reducing sheep infection, and maintenance of refugia. Refugia is a unanimous management practice used for the control of anthelmintic resistance. It consists of retaining part of the parasite population in animals and environment not exposed to anthelmintics, assuring the stock of genes of susceptibility (Kenyon et al., 2009).

Monthly fecal examination was associated with anthelmintic susceptibility, probably because monitoring animal infection can be used as an effective decision criterion for target-selective treatment (Kenyon et al., 2009). Additionally, the criteria used for anthelmintic choice influenced resistance, with beneficial effects observed when the decision was based on technical and efficiency criteria (through FECRT) in contrast to the seller, price, or farmer’s experience.

Farming systems affected resistance, with intensive farming resulting in lower resistance in smaller pasture areas. In intensive systems, higher stock rate and animal concentration may increase contamination. However, better pastures, feeding supplementation, and confinement practices may improve the nutritional conditions of animals, leading to greater resistance to worms (Van Wyk, 1990).

Raising Suffolk sheep was a practice associated with resistance in the classification trees. There are differences in parasite resistance among sheep breeds, and Suffolk shows higher susceptibility to gastrointestinal nematodes than Santa Inês (Amarante et al., 2004). Thus, the selection of more resistant breeds or individuals within breeds is an effective strategy to reduce anthelmintic use in flocks and delay the establishment of resistance (Zvinorova et al., 2016).

Other variables were also identified as important for the prediction of anthelmintic resistance. However, because they were not highlighted in the classification tree nodes, their effects may be distinct based on their interactions with other variables. For example, sheep grazing pastures shared with other animal species have been used for gastrointestinal nematode control. Since most parasites are species-specific, shared grazing with cattle, horses, or pigs can reduce environmental contamination with nematodes (Barger, 1997; Forbes, 2021); however, as an adverse effect, co-grazing sheep with cattle may accelerate the development of anthelmintic resistance, as it may reduce the number of parasites in refugia on pasture (see Falzon et al., 2014 for a review).

Other practices were related to farm structure and technical level, including the number of dams, heads, and sires; total area; cattle grazing; and sheep production as the main income source. These practices reflect farm production levels, care given to sheep, the availability of animals for selection based on parasite susceptibility, and pasture area. Time spent in sheep farming may also have a dual effect; while farming for very short periods could be associated with the use of pastures that are less contaminated with larvae, farming for longer periods may result in learning gains to control the parasites.

Other important practices were associated with anthelmintic treatment protocols. Target-selective and strategic treatments are usually preferred over fixed deworming schedules, but refugia, parasitic load, and animal physical condition should be considered when choosing a deworming protocol (Bath, 2011). Anthelmintic combinations promote nematode control in the presence of single or multiple resistance and slow the development of resistance under certain prerequisites. However, resistance to all drugs used in combination can occur if refugia are insufficient or if resistance is high (reviewed by Bartram et al., 2012). Famacha examinations detect anemia, identify animals for target-selective treatment, and reduce anthelmintic usage, but only when H. contortus, which is hematophagous, is the most prevalent species in the flock (Bath, 2011). Dose-and-move, also known as drench-and-shift, involves moving animals after treatment to a new clean pasture and is considered a high-risk practice for anthelmintic resistance, as it reduces refugia due to contamination of the new pasture only with resistant parasites (Waghorn et al., 2009).

Replacement of the parasite population can occur through the import of animals from other farms; however, the consequences depend on the resistance status of these parasites. Thus, frequent animal incorporation and the region of animal origin can increase anthelmintic resistance when imported helminths are resistant to anthelmintics (Bath, 2011), but the opposite may occur if susceptible worms are imported.

Interestingly, trees provided management alternatives; if a farmer could not modify one practice, it was possible to implement a different strategy to reduce anthelmintic resistance. For example, if changing the forage species of the pasture is impossible, monitoring the fecal egg counts of animals every month can reduce anthelmintic resistance, and the use of intensive farming can compensate the reduced availability of pasture areas in farms. However, as intensification increases not only stocking rate, but also parasitism levels (Waller, 2006), strategies for environmental decontamination and reduction of larvae in pasture, such as using integrated crop-livestock systems (Almeida et al., 2018) or biological control with nematophagous fungi (reviewed by Waller, 2006), may increment the integrated control of gastrointestinal nematodes in sheep flocks.

Classification trees provide a visual interpretation of data and require less computational memory, but are more prone to overfitting and instability, and small perturbations in the dataset can lead to large variations in predictions (Fratello & Tagliaferri, 2019). Due to this disadvantage, a random forest model was also evaluated. Random forest approaches, despite requiring higher computational memory and missing the easy interpretation of the trees, aggregate independent classifiers (or independent trees) and usually result in better performance (Fratello & Tagliaferri, 2019). In the present study, despite selecting almost the same variables of importance, better prediction results were obtained with the classification trees than with the random forest model. Considering the small sample size (27 observations) and unbalanced data (20 resistant and 7 susceptible flocks), building 1,000 classification trees with cross-validation after altering the training set (resampling) resulted in trees with higher prediction accuracy compared to the random forest results, which was based on the aggregation of 1,000 independent trees, but only after one subset of the database.

Conclusions

Machine learning tools are suitable for the fast, easy, noninvasive, and less costly identification and classification of response patterns and their complex relationship to determinant factors. The development of classification trees to predict anthelmintic resistance in sheep flocks based on management practices could provide alternatives for controlling the global multidrug resistance problem. Practices such as the exclusive use of forage species in pastures, monthly FEC examination of animals, intensive farming systems, pasture area available for grazing, and the use of parasite-resistant sheep breeds can be implemented to delay the establishment of anthelmintic resistance in flocks. Furthermore, prioritizing management changes based on variable importance is a smart decision made by farmers and technicians to promote sheep health and nematode control, reduce the use of anthelmintics, and manage the establishment of anthelmintic resistance.

Acknowledgements

Financial support was received from Embrapa (process no. 03.07.09.013.00.00) and FAPESP (processes no. 2021/11842-9 and 2021/02535-5). We appreciate the efforts of the original research team (Adriana Hellmeister de Campos Nogueira, Ana Carolina de Souza Chagas, Ana Lúcia Luz Alberti, Carlos Frederico Carvalho Rodrigues, Cecília José Veríssimo, Cristina Maria Pacheco Barbosa, Daniela Pontes Chiebao, Daniel Cardoso, Giane Serafim da Silva, José Roberto Pereira, Luciana Morita Katiki, Luis Alberto Ambrosio, Luiz Florencio Franco Margatho, Marcelo Beltrão Molento, Marcia Marise de Freitas Cação, Ricardo Lopes Dias da Costa, Romeu Fernandes Nardon, Tatiana Evelyn Hayama Ueno, and Vera Cláudia Lorenzetti Magalhães Curci) in collecting the previously published data.

• How to cite: Niciura SCM, Sanches GM. Machine learning prediction of multiple anthelmintic resistance and gastrointestinal nematode control in sheep flocks. Braz J Vet Parasitol 2024; 33(1): e019023. https://doi.org/10.1590/S1984-29612024014

References

• Almeida FA, Piza MLST, Bassetto CC, Starling RZC, Albuquerque ACA, Protes VM, et al. Infection with gastrointestinal nematodes in lambs in different integrated crop-livestock systems (ICL). Small Rumin Res 2018; 166: 66-72. http://dx.doi.org/10.1016/j.smallrumres.2018.07.009
  » http://dx.doi.org/10.1016/j.smallrumres.2018.07.009
• Amarante AFT, Bricarello PA, Rocha RA, Gennari SM. Resistance of Santa Ines, Suffolk and Ile de France sheep to naturally acquired gastrointestinal nematode infections. Vet Parasitol 2004; 120(1-2): 91-106. http://dx.doi.org/10.1016/j.vetpar.2003.12.004 PMid:15019147.
  » http://dx.doi.org/10.1016/j.vetpar.2003.12.004
• Barger I. Control by management. Vet Parasitol 1997; 72(3-4): 493-500. http://dx.doi.org/10.1016/S0304-4017(97)00113-1 PMid:9460214.
  » http://dx.doi.org/10.1016/S0304-4017(97)00113-1
• Bartram DJ, Leathwick DM, Taylor MA, Geurden T, Maeder SJ. The role of combination anthelmintic formulations in the sustainable control of sheep nematodes. Vet Parasitol 2012; 186(3-4): 151-8. http://dx.doi.org/10.1016/j.vetpar.2011.11.030 PMid:22245073.
  » http://dx.doi.org/10.1016/j.vetpar.2011.11.030
• Bath GF. Non-pharmaceutical control of endoparasitic infections in sheep. Vet Clin North Am Food Anim Pract 2011; 27(1): 157-62. http://dx.doi.org/10.1016/j.cvfa.2010.10.002 PMid:21215899.
  » http://dx.doi.org/10.1016/j.cvfa.2010.10.002
• Bucki M, Dhufaigh KN, O’Brien C, Weatherley A, Walshe N, McElligott T. Comparison of ovine faecal Strongyle egg counts form an accredited laboratory and a rapid, on-site parasite diagnostic system utilising a smartphone app and machine learning. Vet Parasitol 2023; 320: 109976. http://dx.doi.org/10.1016/j.vetpar.2023.109976 PMid:37385104.
  » http://dx.doi.org/10.1016/j.vetpar.2023.109976
• Chagas ACS, Domingues LF, Gaínza YA. Cartilha de vermifugação de ovinos e caprinos São Carlos: Embrapa Pecuária Sudeste; 2013.
• Chagas ACS, Tupy O, Santos IB, Esteves SN. Economic impact of gastrointestinal nematodes in Morada Nova sheep in Brazil. Rev Bras Parasitol Vet 2022; 31(3): e008722. http://dx.doi.org/10.1590/s1984-29612022044 PMid:36000609.
  » http://dx.doi.org/10.1590/s1984-29612022044
• Coles GC. Anthelmintic resistance: looking to the future: a UK perspective. Res Vet Sci 2005; 78(2): 99-108. http://dx.doi.org/10.1016/j.rvsc.2004.09.001 PMid:15563915.
  » http://dx.doi.org/10.1016/j.rvsc.2004.09.001
• Corrêa LA, Santos PM. Manejo e utilização de plantas forrageiras dos gêneros Panicum, Brachiaria e Cynodon São Carlos: Embrapa Pecuária Sudeste; 2009.
• Costa RLD, Bueno MS, Veríssimo CJ, Cunha EA, Santos LE, Oliveira SM, et al. Performance and nematode infection of ewe lambs on intensive rotational grazing with two different cultivars of Panicum maximum. Trop Anim Health Prod 2007; 39(4): 255-63. http://dx.doi.org/10.1007/s11250-007-9005-5 PMid:17847820.
  » http://dx.doi.org/10.1007/s11250-007-9005-5
• Doyle SR, Cotton JA. Genome-wide approaches to investigate anthelmintic resistance. Trends Parasitol 2019; 35(4): 289-301. http://dx.doi.org/10.1016/j.pt.2019.01.004 PMid:30733094.
  » http://dx.doi.org/10.1016/j.pt.2019.01.004
• Falzon LC, O’Neill TJ, Menzies PI, Peregrine AS, Jones-Bitton A, vanLeeuwen J, et al. A systematic review and meta-analysis of factors associated with anthelmintic resistance in sheep. Prev Vet Med 2014; 117(2): 388-402. http://dx.doi.org/10.1016/j.prevetmed.2014.07.003 PMid:25059197.
  » http://dx.doi.org/10.1016/j.prevetmed.2014.07.003
• Food and Agriculture Organization of United Nations - FAO. FAOSTAT [online]. Geneva: FAO; 2022 [cited 2022 Sep 23]. Available from: http://www.fao.org/faostat/en/#data/TP
  » http://www.fao.org/faostat/en/#data/TP
• Forbes AB. Parasites of cattle and sheep: a practical guide to their biology and control Wallingford: CABI; 2021. http://dx.doi.org/10.1079/9781789245158.0000
  » http://dx.doi.org/10.1079/9781789245158.0000
• Fratello M, Tagliaferri R. Decision trees and Random Forest. In: Rnganathan S, Gribskov M, Nakai K, Schonbach C, editors. Encyclopedia of bioinformatics and computational biology: ABC of bioinformatics Amsterdam: Elsevier; 2019. p. 374-83. http://dx.doi.org/10.1016/B978-0-12-809633-8.20337-3
  » http://dx.doi.org/10.1016/B978-0-12-809633-8.20337-3
• Freitas LA, Ferreira REP, Savegnago RP, Dórea JRR, Stafuzza NB, Rosa GJM, et al. Image analysis to automatically classify anemia based on Famacha score in sheep using ocular conjunctiva images. Transl Anim Sci 2023a; 7(1): txad118. http://dx.doi.org/10.1093/tas/txad118
  » http://dx.doi.org/10.1093/tas/txad118
• Freitas LA, Savegnago RP, Alves AAC, Costa RLD, Munari DP, Stafuzza NB, et al. Classification performance of machine learning methods for identifying resistance, resilience, and susceptibility to Haemonchus contortus infections in sheep. Animals 2023b; 13(3): 374. http://dx.doi.org/10.3390/ani13030374 PMid:36766263.
  » http://dx.doi.org/10.3390/ani13030374
• Genuer R, Poggi JM. Random Forests with R Cham: Springer; 2019.
• Greenwell BM. treemisc: data sets and functions to accompany “tree-based methods for statistical learning in R”. R package version 0.0.1 [online]. Vienna: R Foundation for Statistical Computing; 2022 [cited 2023 Nov 16]. Available from: https://CRAN.R-project.org/package=treemisc
  » https://CRAN.R-project.org/package=treemisc
• Instituto Brasileiro de Geografia e Estatística - IBGE. Rebanho de ovinos (ovelhas e carneiros) [online]. Rio de Janeiro: IBGE; 2022 [cited 2022 Sep 23]. Available from: https://www.ibge.gov.br/explica/producao-agropecuaria/ovino/br
  » https://www.ibge.gov.br/explica/producao-agropecuaria/ovino/br
• Izbicki R, Santos TM. Aprendizado de máquina: uma abordagem estatística [online]. São Carlos: UFSCar; 2020 [cited 2023 Nov 16]. Available from: http://www.rizbicki.ufscar.br/AME.pdf
  » http://www.rizbicki.ufscar.br/AME.pdf
• Kenyon F, Greer AW, Coles GC, Cringoli G, Papadopoulos E, Cabaret J, et al. The role of targeted selective treatments in the development of refugia-based approaches to the control of gastrointestinal nematodes of small ruminants. Vet Parasitol 2009; 164(1): 3-11. http://dx.doi.org/10.1016/j.vetpar.2009.04.015 PMid:19450930.
  » http://dx.doi.org/10.1016/j.vetpar.2009.04.015
• Kuhn M. caret: Classification and regression training. R package version 6.0-93 [online]. Vienna: R Foundation for Statistical Computing; 2022 [cited 2023 Nov 16]. Available from: https://CRAN.R-project.org/package=caret
  » https://CRAN.R-project.org/package=caret
• Leathwick DM. Modelling the benefits of a new class of anthelmintic in combination. Vet Parasitol 2012; 186(1-2): 93-100. http://dx.doi.org/10.1016/j.vetpar.2011.11.050 PMid:22169403.
  » http://dx.doi.org/10.1016/j.vetpar.2011.11.050
• Liaw A, Wiener M. Classification and regression by randomForest. R News 2002; 2(3): 18-22.
• McBean D, Nath M, Lambe N, Morgan-Davies C, Kenyon F. Viability of the Happy Factor^TM targeted selective treatment approach on several sheep farms in Scotland. Vet Parasitol 2016; 218: 22-30. http://dx.doi.org/10.1016/j.vetpar.2016.01.008 PMid:26872924.
  » http://dx.doi.org/10.1016/j.vetpar.2016.01.008
• McFarland C, Rose Vineer H, Chesney L, Henry N, Brown C, Airs P, et al. Tracking gastrointestinal nematode risk on cattle farms through pasture contamination mapping. Int J Parasitol 2022; 52(10): 691-703. http://dx.doi.org/10.1016/j.ijpara.2022.07.003 PMid:36113619.
  » http://dx.doi.org/10.1016/j.ijpara.2022.07.003
• McManus C, Louvandini H, Paim TP, Martins RS, Barcellos JOJ, Cardoso C, et al. The challenge of sheep farming in the tropics: aspects related to heat tolerance. R Bras Zootec 2011; 40(Suppl. spe): 107-20.
• McRae KM, Stear MJ, Good B, Keane OM. The host immune response to gastrointestinal nematode infection in sheep. Parasite Immunol 2015; 37(12): 605-13. http://dx.doi.org/10.1111/pim.12290 PMid:26480845.
  » http://dx.doi.org/10.1111/pim.12290
• Miller CM, Waghorn TS, Leathwick DM, Candy PM, Oliver AMB, Watson TG. The production cost of anthelmintic resistance in lambs. Vet Parasitol 2012; 186(3-4): 376-81. http://dx.doi.org/10.1016/j.vetpar.2011.11.063 PMid:22178431.
  » http://dx.doi.org/10.1016/j.vetpar.2011.11.063
• Niciura SCM, Veríssimo CJ, Gromboni JGG, Rocha MIP, Mello SS, Barbosa CMP, et al. F200Y polymorphism in the β-tubulin gene in field isolates of Haemonchus contortus and risk factors of sheep flock management practices related to anthelmintic resistance. Vet Parasitol 2012; 190(3-4): 608-12. http://dx.doi.org/10.1016/j.vetpar.2012.07.016 PMid:22858226.
  » http://dx.doi.org/10.1016/j.vetpar.2012.07.016
• Oliveira ALF, Costa C, Rodella RA, Silva BF, Amarante AFT. Effect of plant trichomes on the vertical migration of Haemonchus contortus infective larvae on five tropical forages. Trop Anim Health Prod 2009; 41(5): 775-82. http://dx.doi.org/10.1007/s11250-008-9251-1 PMid:18975119.
  » http://dx.doi.org/10.1007/s11250-008-9251-1
• Pires CRS, Cavalcante GG, Prado RS, Costa LM Jr, Rivero GRC. Survival of Haemonchus contortus larvae in forage species in the Eastern Amazon. Pesq Agropec Bras 2021; 56: e02664. https://doi.org/10.1590/S1678-3921.pab2021.v56.02664
  » https://doi.org/10.1590/S1678-3921.pab2021.v56.02664
• R Core Team. R: a language and environment for statistical computing [online]. Vienna: R Foundation for Statistical Computing; 2022 [cited 2023 Nov 16]. Available from: https://www.R-project.org/.2022
  » https://www.R-project.org/.2022
• Santos MC, Silva BF, Amarante AFT. Environmental factors influencing the transmission of Haemonchus contortus. Vet Parasitol 2012; 188(3-4): 277-84. http://dx.doi.org/10.1016/j.vetpar.2012.03.056 PMid:22521972.
  » http://dx.doi.org/10.1016/j.vetpar.2012.03.056
• Sargison ND. Pharmaceutical treatments of gastrointestinal nematode infections of sheep: future of anthelmintic drugs. Vet Parasitol 2012; 189(1): 79-84. http://dx.doi.org/10.1016/j.vetpar.2012.03.035 PMid:22497871.
  » http://dx.doi.org/10.1016/j.vetpar.2012.03.035
• Speybroeck N. Classification and regression trees. Int J Public Health 2012; 57(1): 243-6. http://dx.doi.org/10.1007/s00038-011-0315-z PMid:22015650.
  » http://dx.doi.org/10.1007/s00038-011-0315-z
• Steinberg D. CART: classification and regression trees. In: Wu X, Kumar V, editors. The top ten algorithms in data mining 1^st ed. New York: Taylor & Francis Group; 2009. p. 179-201. http://dx.doi.org/10.1201/9781420089653.ch10
  » http://dx.doi.org/10.1201/9781420089653.ch10
• Suresh KP, Sengupta PP, Jacob SS, Sathyanarayana MKG, Patil SS, Swarnkar CP, et al. Exploration of machine learning models to predict the environmental and remote sensing risk factors of haemonchosis in sheep flocks of Rajasthan, India. Acta Trop 2022; 233: 106542. http://dx.doi.org/10.1016/j.actatropica.2022.106542 PMid:35643184.
  » http://dx.doi.org/10.1016/j.actatropica.2022.106542
• Therneau T, Atkinson B. rpart: recursive partitioning and regression trees. R package version 4.1.16 [online]. 2022 [cited 2023 Nov 16]. Available from: https://CRAN.R-project.org/package=rpart
  » https://CRAN.R-project.org/package=rpart
• Ueno H, Gonçalves PC. Manual para diagnóstico das helmintoses de ruminantes 4. ed. Tokyo: Japan International Cooperation Agency; 1998.
• Urzedo MCC, Santos MER, Cuccato LP, Macedo SAS, Rosalinski-Moraes F. Vertical migration of Haemonchus sp. infectant larvae on Stylosanthes spp. and Brachiaria brizantha (Syn. Urochloa brizantha) cv. marandu forage. Biosci J 2022; 38: e38060. http://dx.doi.org/10.14393/BJ-v38n0a2022-53801
  » http://dx.doi.org/10.14393/BJ-v38n0a2022-53801
• Van Wyk JA. Integrated worm control as a strategy in the control of gastrointestinal nematodes of sheep and cattle. J S Afr Vet Assoc 1990; 61(3): 141-5. PMid:2287005.
• Veríssimo CJ, Niciura SCM, Alberti ALL, Rodrigues CFC, Barbosa CM, Chiebao DP, et al. Multidrug and multispecies resistance in sheep flocks from São Paulo state, Brazil. Vet Parasitol 2012; 187(1-2): 209-16. http://dx.doi.org/10.1016/j.vetpar.2012.01.013 PMid:22341829.
  » http://dx.doi.org/10.1016/j.vetpar.2012.01.013
• Voigt K, Geiger M, Jäger MC, Knubben-Schweizer G, Strube C, Zablotski Y. Effectiveness of anthelmintic treatments in small ruminants in Germany. Animals (Basel) 2022; 12(12): 1501. http://dx.doi.org/10.3390/ani12121501 PMid:35739838.
  » http://dx.doi.org/10.3390/ani12121501
• Waghorn TS, Miller CM, Oliver A-MB, Leathwick DM. Drench-and shift is a high-risk practice in the absence of refugia. N Z Vet J 2009; 57(6): 359-63. http://dx.doi.org/10.1080/00480169.2009.64723 PMid:19966896.
  » http://dx.doi.org/10.1080/00480169.2009.64723
• Walkden-Brown SW, Kahn LP. Nutritional modulation of resistance and resilience to gastrointestinal nematode infection: a review. Asian-Australas J Anim Sci 2002; 15(6): 912-24. http://dx.doi.org/10.5713/ajas.2002.912
  » http://dx.doi.org/10.5713/ajas.2002.912
• Waller PJ. Management and control of nematode parasites of small ruminants in the face of total anthelmintic failure. Trop Biomed 2004; 21(2): 7-13. PMid:16493393.
• Waller PJ. Sustainable nematode parasite control strategies for ruminant livestock by grazing management and biological control. Anim Feed Sci Technol 2006; 126(3-4): 277-89. http://dx.doi.org/10.1016/j.anifeedsci.2005.08.007
  » http://dx.doi.org/10.1016/j.anifeedsci.2005.08.007
• Zamanian M, Chan JD. High-content approaches to anthelmintic drug screening. Trends Parasitol 2021; 37(9): 780-9. http://dx.doi.org/10.1016/j.pt.2021.05.004 PMid:34092518.
  » http://dx.doi.org/10.1016/j.pt.2021.05.004
• Zvinorova PI, Halimani TE, Muchadeyi FC, Matika O, Riggio V, Dzama K. Breeding for resistance to gastrointestinal nematodes: the potential in low-input/output small ruminant production systems. Vet Parasitol 2016; 225: 19-28. http://dx.doi.org/10.1016/j.vetpar.2016.05.015 PMid:27369571.
  » http://dx.doi.org/10.1016/j.vetpar.2016.05.015

Publication Dates

• Publication in this collection
  18 Mar 2024
• Date of issue
  2024

History

• Received
  16 Nov 2023
• Accepted
  18 Jan 2024