PS10-27. Integrating Multi-Frequency Bioelectrical Impedance and Machine Learning for Pregnancy Detection in Sheep.

Abstract

Early pregnancy detection is critical for improving reproductive management and production efficiency in sheep. While conventional diagnostic methods such as ultrasonography and blood-based assays are reliable, they are often resource-intensive and limit frequent on-farm monitoring. Therefore, there is a growing interest in non-invasive and repeatable approaches for continuous physiological assessment. Bioelectrical impedance analysis (BIA) is a rapid, non-invasive technique that measures tissue electrical properties and may reflect physiological changes associated with pregnancy, including alterations in tissue composition, fluid distribution, and electrolyte balance. This study aimed at evaluating the association between multi-frequency BIA measurements and pregnancy status in sheep and to assess the predictive potential of impedance-derived features using machine learning (ML) approaches. Thirty-six synchronized crossbred ewes were naturally bred using two mature (2-3 year) Katahdin rams and pregnancy was confirmed at 28 days post-breeding (PB) using transabdominal ultrasonography. Bioelectrical impedance measurements, including resistance (Rs), reactance (Xc), impedance magnitude (Z), and phase angle (PA), were collected at 10, 50,100, and 180 kHz frequencies from vaginal and groin regions at multiple time points (pre-breeding, and at 28, 50, and 100 days PB in pregnant ewes). Data were analyzed with mixed model analysis and Tukey’s means comparison. In groin region, Rs (Ω), Xc (Ω), Z (Ω), and PA (°) values were higher (P < 0.05) in pregnant ewes at 100-day PB than before breeding and 28-day PB at all frequencies. However, in the vaginal region these values decreased at 100-day PB than other time points at all frequencies. To further investigate the predictive potential of these features, various ML models, including Gradient Boosting (GB), Backpropagation Neural Network–Multi-Layer Perceptron (BPNNMLP), Support Vector Machine with Radial Basis Function Kernel (SVMRBF), Random Forest (RF), K-Nearest Neighbors (KNN), and Logistic Regression (LR) were implemented. Model performance demonstrated moderate classification ability, with GB achieving the highest accuracy (0.62 ± 0.20) compared to other models (ranging from 0.42 to 0.57). In conclusion, these findings suggest that multi-frequency BIA can be a potential non-invasive tool for capturing biological variations associated with pregnancy in sheep. The integration of BIA with ML provides a promising framework for developing data-driven approach to support reproductive monitoring and precision livestock management.