PS11-3. A Digital Twin Framework For High-risk FAMACHA Prediction in Goats Using Bioelectrical Impedance Analysis and Grouped Machine Learning.

Abstract Early identification of anemia risk in goats is important for improving health monitoring, parasite control, and decision-making at the animal level. This study developed a digital twin-oriented machine learning framework to predict FAMACHA-derived anemia risk using bioelectrical impedance analysis (BIA) and contextual environmental variables while preserving repeated-measures independence through grouped cross-validation by animal. An analysis-ready dataset containing 183 observations from 81 goats collected across 6 sampling days was used. Objective FAMACHA score validated using appropriate blood packed cell volume ranges served as the primary reference label, and field FAMACHA scores were retained for agreement analysis and practical relevance. Predictor variables included resistance and reactance features measured at 50, 100, and 180 kHz, derived impedance ratios, deltas and slopes across frequencies, fecal egg count, and day-level weather covariates. For practical deployment, binary classification was defined as low risk (FAMACHA 1–2) versus high risk (FAMACHA 3–4), while 4-class and collapsed 3-class models were also evaluated as supportive analyses. The dataset showed class imbalance, with most samples in objective classes 1 and 2 and relatively few in classes 3 and 4. Agreement between field and objective FAMACHA was high, with 94.5% exact agreement and a quadratic weighted kappa of 0.934, indicating strong alignment between scoring methods. Among binary grouped machine learning models, logistic regression provided the best overall practical performance, achieving a balanced accuracy of 0.754, macro F1 of 0.691, ROC AUC of 0.872, sensitivity of 0.667 for high-risk animals, and specificity of 0.841 for low-risk animals. Random forest produced similar ROC AUC but substantially lower sensitivity and macro F1, indicating weaker balance in identifying high-risk animals. Multiclass performance was lower than binary performance, with best macro F1 values of 0.378 for the 4-class model and 0.473 for the collapsed 3-class model, suggesting that binary high-risk prediction is the more reliable operational endpoint under the current data structure.Out-of-fold probabilities from the best binary model were converted into digital twin risk states, allowing each animal-day-site observation to be assigned as low, moderate, or high risk. Feature-weight analysis indicated that prediction was driven by a combination of impedance variability, derived frequency features, fecal egg count, site effects, and environmental context. These results support the use of BIA-based, animal-level digital twin modeling as a promising framework for longitudinal anemia-risk monitoring and targeted decision support in goats.