Evaluating the Efficacy of Bioelectrical Impedance Analysis Using Machine Learning Models for the classification of Parasitized Goats

1. Introduction

In warm and wet regions, small ruminants such as goats and sheep are especially vulnerable to parasitic infections, with Haemonchus contortus posing one of the greatest threats [1]. Commonly referred to as the barber pole worm due to its distinctive appearance, Haemonchus contortus thrives in warm, humid environments, infecting the abomasum (the fourth stomach chamber) of ruminants [2,3] by attaching to stomach lining and feeds on the host’s blood, leading to substantial blood loss. Over time, infected animals develop anemia, which severely impacts their overall health, productivity, and even survival [2]. For farmers and livestock producers, infections from Haemonchus contortus represent a significant economic burden, causing diminished growth rates, lowered reproductive success, reduced milk production, and, in severe infestations, death.

Anemia in infected animals is a hallmark sign of Haemonchus contortus infection and is traditionally diagnosed using Hematocrit Analysis, a method that measures the packed cell volume (PCV) of blood [4,5]. The PCV analysis provides an estimate of the level of anemia, reflecting the degree to which the animal is suffering from blood loss. Although reliable, Hematocrit Analysis is labor-intensive, requiring invasive blood sampling and the expertise of trained personnel. This limits its scalability, especially for large herds, and introduces additional stress to the animals [6,7]. The cost and time involved in performing Hematocrit Analysis make it less practical for routine monitoring in large or resource-limited operations.

Another traditional approach used to diagnose parasitic infections in small ruminants is the fecal egg count (FEC). This method involves examining the number of parasite eggs present in an animal’s feces to estimate the parasite burden within the host [8,9]. Fecal egg count is particularly useful for monitoring gastrointestinal nematodes such as Haemonchus contortus, as the number of eggs passed in the feces correlates with the worm burden in the animal [10,11]. The FEC method is commonly used in farm management programs to assess the need for treatment and to monitor the efficacy of deworming programs [12,13]. However, like Hematocrit Analysis, FEC is labor-intensive, requiring specialized equipment and expertise to collect, process, and interpret fecal samples [14,15]. Additionally, FEC results can vary significantly depending on the animal's diet, hydration status, and the time of day when samples are collected. Moreover, the accuracy of fecal egg counts may decrease when the parasite burden is low, leading to potential underestimation of the infection [16].

To address the limitations of traditional diagnostic methods such as PCV and FEC, researchers have been exploring alternative technologies that offer rapid, non-invasive, and scalable solutions. One such promising tool is bioelectrical impedance devices for analysis (BIA), a technique that measures the electrical properties of biological tissues, specifically electrical resistance (R) and electrical reactance (Xc) [17,18]. BIA works by passing a small, painless electrical impulse through the body and measuring how the tissues oppose the flow of the current [19,20]. These measurements taken through BIA provide insights into the body’s composition, including water content, fat mass, and cellular health [21,22], all of which can be affected by parasitic infections.

In the case of Haemonchus contortus, the parasite's blood-feeding behavior disrupts the host’s fluid balance and reduces the total volume of blood and red blood cells, which can alter the electrical properties of the tissue. This makes BIA a promising tool for detecting parasitic infections, as changes in the body’s fluid levels and tissue composition are directly related to the severity of the infection. Unlike Hematocrit Analysis or FEC, BIA can be performed quickly, without the need for invasive procedures or specialized personnel. This makes BIA particularly well-suited for large-scale herd management, where rapid, non-invasive diagnostics are crucial for maintaining animal health and productivity.

Given the advantages of BIA in terms of ease, speed, and non-invasiveness, this study explored its potential in combination with machine learning (ML) techniques to classify goats as either parasitized or healthy. Machine learning algorithms, known for their ability to recognize complex patterns in large datasets, are applied to the bioelectrical impedance data collected from live goats. By training machine learning models on this data, the study aims to develop an efficient, scalable, and accurate diagnostic tool for detecting parasitic infections in small ruminants. The implications of this research go beyond the detection of Haemonchus contortus. The successful application of BIA and machine learning could pave the way for diagnosing a wide range of health conditions in livestock, from other parasitic infections to metabolic disorders and nutritional deficiencies. By providing a rapid, non-invasive, and cost-effective diagnostic tool, this study lays the groundwork for the integration of advanced technologies in precision livestock farming, where early detection and proactive management are key to maximizing productivity and animal welfare. The following sections 1.1, 1.2, 1.3, 1.4 provides the basic working principle for the different techniques used in the study.

1.1. Support Vector Machines (SVM)

Support Vector Machines (SVMs) are a category of supervised machine learning methods employed for classification, regression, and anomaly detection. The SVM technique is widely recognized in machine learning for its efficacy in handling both linear and nonlinear classification challenges. The fundamental concept of SVM is to identify a hyperplane in an N-dimensional space (where N denotes the number of features) that effectively segregates data points into various classes [23,24]. Winston [25] compares this approach to “Fitting the widest possible street,” elucidating the quadratic optimization challenge associated with hyperplane separation using the equation [26]:

wx + b = 0

(1)

w represents the weight vector, x is the input vector, and b signifies the bias factor. In SVM, hard-margin and soft-margin classifiers identify the optimal separation distances between classes, facilitating the classification of distinct patterns or features from an image [27].

In binary classification, numerous hyperplanes can delineate the data points of two classes, whereas in multi-class classification, more than two classes are taken into account. The goal is to identify the hyperplane that maximizes the margin, defined as the maximum distance between the support vectors data points nearest to the hyperplane. The support vectors are essential, as their removal would change the hyperplane's position. Consequently, SVMs are regarded as essential elements of the dataset. The efficacy of SVM resides in its capacity to manage nonlinear data using the kernel trick, which converts non-linearly separable data into a linearly separable format [28,29,30].

1.2. K-Nearest Neighbor (K-NN)

K-Nearest Neighbors (KNN) is a basic and efficient non-parametric, supervised machine learning classifier which utilizes proximity to categorize or forecast the group to which a data item is allocated. The method operates by retaining all existing data points and categorizing new instances according to their similarity, generally assessed by distance metrics such as Euclidean, Manhattan, Minkowski, or Hamming distances [31]. The primary objective of KNN is to determine the nearest neighbors to a specified query point and subsequently allocate a class label. These distance functions facilitate the evaluation of proximate points, with the ultimate categorization established using a majority vote mechanism among the nearest neighbors. Each data point is categorized according to the predominant category among its closest neighbors, as established by the selected distance metric. The selection of K substantially influences prediction accuracy, lower values render the model susceptible to noise, whereas higher values elevate processing requirements. In datasets comprising two classes, researchers frequently select an odd value for K to prevent ties. Nonetheless, a limitation of KNN is that its processing performance may significantly diminish as the dataset size increases [30,31,32,33].

1.3. Back Propagation Neural Networks (BPNN)

Back Propagation Neural Networks (BPNNs) are the part of Artificial Neural Networks (ANNs) that employ the backpropagation method for training purposes. They are well acknowledged for their efficacy in deep learning models. A BPNN comprises a minimum of three layers of nodes: an input layer, one or more hidden layers, and an output layer. Every node, known as an artificial neuron or perceptron, that is interconnected by weighted linkages, which are modified throughout the training process [28,30,34,35]. Multilayer perceptrons employ the backpropagation method, consisting of two phases: a forward pass, in which input data is processed through the network to produce an output, and a backward pass, during which the error (the differences between the anticipated and actually output) flows backward through the network to adjust the weights. This modification is executed utilizing optimization techniques, including gradient descent. Backpropagation Neural Networks (BPNNs) are advantageous due to their capacity to learn and represent complicated non-linear relationships between values[30,35]. Upon completion of training, algorithms may generate precise predictions when presented with novel data, rendering them exceptionally adaptable and versatile for real life usage. For a comprehensive elucidation of BPNN functionality, refer to Siddique et al. [29]. These models demonstrate proficiency in tasks necessitating pattern recognition and are trained utilizing features derived from diverse datasets, such images, numerical data, or text [36].

1.4. Extreme Gradient Boosting (XGBoost)

Extreme Gradient Boosting (XGBoost) is a complex delicate machine learning technique derived from the gradient boosting framework. The fundamental idea is to enhance forecast performance by incrementally constructing an ensemble of decision trees, with each tree rectifying the weaknesses of its predecessors. In the training process, XGBoost initially establishes a rudimentary model (often a constant value) and subsequently computes the residuals or discrepancies between the predicted and actual values. In succeeding steps, a new decision tree is fitted to minimize these residuals using gradient descent, a method that modifies the model by pursuing the direction of maximal error reduction. New trees are incorporated into the ensemble to forecast the residuals, while the model adjusts the weights of misclassified examples to mitigate subsequent errors [37].

XGBoost integrates various distinctive optimizations. It employs regularization approaches (L1 and L2) to mitigate overfitting, hence assuring the model generalizes effectively to novel data [38]. The algorithm adeptly manages absent values and sparse datasets, autonomously identifying the optimal trajectory through the decision tree in the absence of data. XGBoost also executes tree pruning, ceasing tree growth when additional splits yield negligible enhancements, so improving both performance and efficiency [36]. XGBoost is esteemed for its scalability and quickness due to its capacity for parallel data processing, rendering it an optimal selection for managing extensive datasets and intricate issues. The aforementioned qualities have rendered XGBoost highly esteemed in both academic research and industrial applications, particularly in regression and classification tasks [36,37].

2. Materials and Methods

Study Design

A total of 94 intact male Spanish goats (58 healthy; 36 parasitized) (24 months old; 36-50 kg) were used in this experiment over multiple time points from August 2023 to October 2023, providing a robust dataset for analysis. This longitudinal approach allowed us to capture variations over time while adhering to ethical guidelines. All animal use protocols were approved by the Fort Valley State University (FVSU, Fort Valley, GA, USA) Agricultural and Laboratory Animal Care and Use Committee (ALACUC approval number F-T-01-2022). The study was carried out especially to ensure that regulations regarding animal welfare were followed by reducing any conditions that would cause the experimental goats' deaths, mainly caused by low PCV values. The goats were allowed to acquire a natural parasitic infection by grazing on grass pasture at the FVSU Agriculture Technology Center farm from March through September 2023. Healthy goats were monitored weekly for parasites and regularly dewormed using commercially available dewormers with veterinarian recommended doses. Several machines learning models, including Support Vector Machines (SVM), Back Propagation Neural Network, k-Nearest Neighbors (k-NN), XGBoost, and Keras deep learning models were trained and evaluated to determine their ability to classify goats based on their bioelectrical properties especially electrical resistance and electrical reactance. This study also aimed to assess the variations in bioelectrical impedance parameters such as electrical resistance (Rs) and electrical reactance (Xc) between healthy and parasitized goats. A total of 1,540 observations points were collected from August 2023 to December 2023, consisting of 917 healthy goat observation points and 623 parasitized goats observation points from 94 goats throughout the experimental period. Bioelectrical impedance analysis (BIA) was employed to assess Rs and Xc, which reflect tissue hydration levels and cellular integrity.

3. Data Collection and Preprocessing

Data collection was conducted using an online cloud-based server specifically designed by the BIA device provider, which was preprocessed by dividing it into training and testing sets to assess model performance. For classification and regression tasks, features were scaled using standard techniques like normalization to ensure uniformity across the dataset [39]. This step was critical, especially for machine learning algorithms like Support Vector Machines (SVM) and K-Nearest Neighbors (KNN), which are sensitive to feature scaling [40,41]. The target variable ‘y_test’ was adjusted based on the problem type: classification models predicted categorical labels, while regression models predicted continuous values [42]. A 10-fold nested cross-validation approach was employed to reduce overfitting and evaluate model performance comprehensively (Varma & Simon, 2006). This method included an inner loop for hyperparameter optimization and an outer loop for model evaluation, thus providing more reliable performance estimates [43,44].

The statistical analyses were also conducted with PROC GLIMMIX in SAS to evaluate the impact of the condition (healthy versus parasitized) on Rs and Xc. The Poisson distribution with a logarithmic link function was employed to describe the response variables based on the characteristics of the data. Least squares mean (LS Means) were computed for each condition, and Tukey-Kramer corrections were applied to address multiple comparisons. Furthermore, PROC MEANS was employed to compute descriptive statistics, encompassing the mean, standard deviation, and standard error for each variable.

3.1. Model Development and Pipeline

The model development process comprised two primary components: classification and regression tasks, examined by several machine learning methods.

3.1.1. Classification Models Pipeline

Backpropagation Neural Networks (BPNN), SVM, and KNN, were used for classification based analysis, each of which was evaluated using accuracy and AUC-ROC (Area Under the Receiver Operating Characteristic Curve) scores (Bishop, 2006). The ‘classification_summary’ function executed the process of predicting the test set outcomes and calculating the respective performance metrics [45]. The accuracy was computed using the ‘accuracy_score’ function from the scikit-learn library [45,46], which provides a direct comparison between the predicted and actual labels [47]. For models capable of generating probability estimates (e.g., SVM and BPNN), the AUC was calculated using the ‘roc_auc_score’ function, providing insight into how well the model distinguishes between different classes [48]. For models without probabilistic outputs (like KNN), the AUC score was marked as 'N/A' [42]. The results for each model were compiled into a panda ‘DataFrame’ to allow for the comparison of classification accuracy and AUC values [49].

3.1.2. Regression Models Pipeline

The regression pipeline incorporated modified models for regression based analysis which includes BPNN, SVM, KNN, XGBoost, and a Keras-based Neural Network [37,50]. The ‘regression_summary’ function evaluated these models using two primary performance metrics: the R² score and Mean Squared Error (MSE). The R² score measured the proportion of variance in the test data explained by the model [51], while MSE quantified the average squared differences between the predicted and actual values, highlighting the overall prediction error [52]. This pipeline facilitated the comparison of model performance by storing these results in a ‘pandas DataFrame’, providing a clear view of the most effective regression algorithms for the given dataset [49]. The integration of a Keras neural network was particularly useful in exploring the capabilities of deep learning approaches in solving regression tasks [53].

3.1.3. Nested Cross-Validation

A 10-fold nested cross-validation method was employed to enhance model resilience and mitigate the possibility of overfitting [43,54]. In the outer loop, the dataset comprising 1540 observation points was partitioned into 10 equal folds, with each fold utilized as the test group once, while the remaining nine folds or groups were employed for training. Hyperparameter optimization for the inner loop was conducted using an additional 10-fold cross-validation within the training set, ensuring the model's parameters were refined without incorporating test data into the training phase [43]. This nested technique yielded more dependable performance estimates by considering both model selection and model evaluation [42,44].

The layered cross-validation technique facilitated thorough optimization of model hyperparameters, especially for models like SVM and XGBoost, which necessitate meticulous parameter selection for peak performance [37,55]. The inner cross-validation loop guaranteed that model selection relied on optimal parameters derived from training data, without affecting the model's performance assessment in the outer loop [54]. This validation method was essential for both classification and regression processes, guaranteeing that the final performance measures appropriately represented the model's generalization capacity [44]. The performance measures from all outer folds were summed to yield a reliable assessment of the model's efficacy throughout the complete dataset [43]. This complex validation method produced more dependable models, especially in instances where the used dataset demonstrated variability [56]. It also improves the confidence in model predictions for both classification and regression tasks, rendering it a significant element of our investigation [44,50,56].

4. Results and Discussion

The PROC GLIMMIX method revealed a significant impact of condition (healthy vs. parasitized) on electrical resistance (Rs) (F = 635.36, P < 0.0001). Parasitized goats demonstrated a significantly higher least squares mean electrical resistance (5.5321 ± 0.002520 SE) in comparison to healthy goats (5.4483 ± 0.002166 SE). The difference between the two categories was significantly different (Tukey-adjusted P < 0.0001), with parasitized goats exhibiting an increase of 0.08377 units in electrical resistance. The PROC MEANS approach yielded descriptive statistics, revealing that the mean Rs for healthy goats was 232.37 ± 5.77 SE, but parasitized goats demonstrated an average mean Rs of 252.68 ± 7.31 SE, hence corroborating the enhanced electrical resistance in parasitized animals.

This study's results indicate that parasite infections markedly affect the bioelectrical impedance parameters of goats, especially electrical resistance (Rs). Parasitized goats exhibited a significant elevation in Rs relative to healthy goats, potentially attributed to blood loss, dehydration, and tissue changes induced due to the GIN parasite infection [57]. The parasite feeds on the host's blood, resulting in anemia and lower fluid volume, hence elevating tissue electrical resistance when electrical current encounters increased opposition in drier, less hydrated tissues [1,58,59].

A notable difference in electrical reactance (Xc) was also observed between healthy and parasitized goats (F = 11.12, P = 0.0009). Parasitized goats had greater least squares mean electrical reactance (3.9628 ± 0.005524 SE) than healthy goats (3.9388 ± 0.004608 SE), although the disparity was less pronounced than that observed for electrical resistance. The Tukey-adjusted comparison indicated statistical significance (P = 0.0009), with a mean difference of 0.02399 units in electrical reactance. Descriptive statistics using PROC MEANS indicated mean Xc values of 51.36 ± 4.26 SE for healthy goats and 52.60 ± 5.23 SE for parasitized goats. Although the difference in electrical reactance was statistically significant, the size of the change was lesser than that of electrical resistance, suggesting a less prominent effect.

The rise in electrical reactance (Xc) in parasitized goats was statistically significant, although the impact was lower to that of electrical resistance (Table 1). Electrical reactance indicates cell membrane integrity and fluid distribution, and the slight rise in Xc implies that parasite infections exert a limited influence on the capacitive characteristics of tissues [60]. This may result from tissue injury and cell membrane disruption, which influence the storage of electrical current in the tissues, however the impact is less significant than electrical resistance [61,62,63,64]

Table 2 summarizes the classification performance of the models, presenting the Accuracy, Precision, Recall, and F1-Score for each model. Among the models, the SVM exhibited superior performance, achieving an accuracy of 95%, a precision of 93%, and an F1-Score of 96%, demonstrating its robust capacity to differentiate between healthy and parasitized goats. XGBoost achieved an accuracy of 94% and an F1-Score of 95%, indicating the efficacy of ensemble approaches in managing intricate datasets [37,50]. The BPNN model demonstrated a high level of performance, with an accuracy of 92% and an F1-Score of 94%, positioning it as a competitive alternative for classification tasks [65,66]. Keras DL attained an accuracy of 91% and an F1-Score of 93%, indicating that deep learning models can perform effectively, although they were unable to exceed SVM or XGBoost [67,68]. Ultimately, K-NN attained the lowest accuracy of 90%, suggesting it underperformed relative to more advanced models such as SVM and XGBoost, perhaps because to its simplicity and susceptibility to data noise and non-linearity in data [67].

The robust efficacy of SVM and XGBoost in classification tasks can be linked to several factors. Support Vector Machine (SVM) operates by identifying the ideal hyperplane that maximizes the margin between classes, rendering it particularly successful for linearly separable datasets [42,50,55,69]. Considering that bioelectrical impedance measurements probably demonstrate unique patterns between healthy and parasitized goats, the capability of SVM to establish a definitive separation in the feature space enables it to attain high accuracy. Furthermore, SVM excels with high-dimensional data, potentially demonstrating its higher efficacy compared to simpler models such as K-NN [33,42].

XGBoost demonstrated excellent performance, which is anticipated due to its capacity to manage intricate, non-linear interactions via boosting. XGBoost captures complex patterns in the dataset by systematically rectifying errors from prior iterations [37,50]. This is especially advantageous in biological datasets because nuanced variations in characteristics can significantly influence classification [42,70]. The BPNN and Keras deep learning models exhibited marginally reduced performance compared to SVM and XGBoost, although they still demonstrated competitive outcomes [65]. Neural networks probably encapsulate intricate, non-linear relationships within the data; yet their efficacy may be affected by the selection of hyperparameters or the dimensions of the neural network. The K-NN method, due to its reliance on proximity-based judgments, had difficulties managing the dataset's complexity, maybe accounting for its lower accuracy relative to other models [51,68].

Table 3 shows the efficacy of the models in forecasting the extent of parasitism, as indicated by the R-squared (R²) value and Mean Squared Error (MSE). The BPNN model demonstrated superior performance, achieving a R² value of 99.9% and a minimal MSE of 1.25e-04, signifying its exceptional predictive accuracy about the health status of goats [65]. SVR demonstrated strong performance, achieving a R² value of 96.9% and a minimal MSE of 7.69e-03, indicating its high reliability as a regression model [37,50]. XGBoost and Keras DL attained R² values of 89.2% and 88%, respectively, with moderate MSE values, suggesting that although these models exhibited better performance, they lacked the precision of BPNN or SVR. K-NN regression exhibited the poorest performance, with a R² value of 83% and a higher MSE of 3.30e-02, indicating its worse ability to describe the intricate correlations between bioelectrical data and goat health problems relative to the other models [33].

The BPNN regression model was the most successful model in regression tasks, obtaining a nearly perfect R² value of 99.9%. This suggests that the neural network model was able to accurately anticipate the degree of parasitism by capturing nearly all of the variance in the dataset [65,68]. The success of BPNN in regression can be attributed to its capacity to learn complex, non-linear patterns, which is particularly advantageous in biological data that involve interactions between various physiological parameters. BPNN outperformed other models in regression tasks because neural networks are well-suited for encoding these intricate relationships [71].

Additionally, SVR demonstrated better performance, attaining a R² value of 96.9%. SVR operates similarly to SVM, but it is designed to identify a function that deviates from the true data points by a small margin for continuous data [72]. The efficacy of SVR in predicting parasitism severity may have been influenced by its capacity to manage outliers and noisy data. XGBoost and Keras DL also demonstrated strong predictive potential; however, they were unable to achieve the same level of precision as BPNN or SVR. This could be attributed to the hyperparameter tuning limitations or the complexity of the data [37,51]. However, K-NN regression encountered difficulty with this task, most likely due to its dependence on local averaging, which may not adequately capture the global trends apparent in the dataset, as opposed to more advanced models such as BPNN and SVR [32,68].

5. Conclusions

Bio-electrical impedance analysis approach has the potential to revolutionize parasitism detection and livestock management practices. Traditional methods like FEC and Hematocrit Analysis, while effective, require substantial labor, time, and expertise, making them less practical for large-scale or resource-constrained operations. In contrast, BIA, combined with machine learning, can provide real-time diagnostic information without the need for invasive procedures or specialized personnel. This could significantly enhance the ability of farmers and veterinarians to monitor the health of their herds, enabling earlier detection and treatment of parasitic infections, reducing production losses, and improving animal welfare.

This study shows the importance of bioelectrical impedance analysis (BIA) in goat parasitism diagnosis. Machine learning models effectively classified goats as healthy or parasitized based on bioelectrical impedance. Support Vector Machine (SVM) and Backpropagation Neural Networks (BPNN) were the most successful models, with BPNN obtaining near-perfect accuracy in predicting parasitism (R² value of 99.9%). BPNN's capacity to capture complicated, non-linear biological data correlations is vital for dealing with physiological fluctuations in parasitized animals, explaining its excellent performance. Parasitic infections alter goat tissue electrical characteristics, as seen by differences in electrical resistance (Rs) and electrical reactance (Xc) between healthy and parasitized goats. Due to dehydration and blood loss from parasitic diseases like Haemonchus contortus, parasitized goats had increased electrical resistance (Rs). These illnesses cause anemia, which reduces blood volume and tissue fluid retention, raising electrical current electrical resistance. This supports prior studies indicating that parasite infections in small ruminants alter hydration and tissue composition, which bioelectrical impedance may detect. High classification accuracy of machine learning models like SVM and XGBoost shows their robustness in high-dimensional, complicated datasets like bioelectrical impedance measurements. These models are ideal for real-time diagnostics because they can detect minor differences between healthy and parasitized animals.

Future Research

Future research can build on this study in multiple ways. First, bioelectrical impedance analysis (BIA) can be used to diagnose more animal health issues. BIA can detect tissue changes caused by metabolic abnormalities, dietary deficits, and chronic illnesses. Studying BIA in these circumstances could yield useful insights and non-invasive ways for early diagnosis of livestock (including ruminants and small ruminants) health concerns. Further research into deep learning models may improve livestock examinations machine learning prediction. Test more advanced neural networks like CNNs or RNNs to determine whether they improve the present models. When paired with larger datasets or time-series data from repeated BIA tests, these models may better capture temporal trends and long-term dependencies. Researchers could construct algorithms that detect and forecast parasitism by tracking goats and collecting BIA data at different phases of infection. Early intervention and improved herd health and productivity may result from more effective treatment regimens. These methods could also be applied to sheep, and cattle to test the generalizability of BIA and machine learning models across farming systems. Understanding how parasitism or other health issues affect each species' bioelectrical characteristics can help build species-specific diagnostic tools. Finally, future research should also examine the economic benefits of BIA and machine learning for health issues identification. Early infection prevention can save treatment costs, productivity losses, and animal welfare, improving farm profitability and sustainability.