Phenomics

Utility of multi-modal integration for predicting postpartum diseases in dairy cattle with imbalanced data

---

Abstract

Early detection of postpartum diseases in dairy cattle is essential for health, welfare, and productivity, but current methods rely on visual monitoring and clinical evaluations, often delaying diagnosis. Integrating time-series behavioral sensor data with single time-point management records provides new opportunities for individualized and timely disease detection, but requires advanced multimodal modeling approaches. In addition, disease cases are often much fewer than healthy controls, creating imbalanced data that limits model prediction accuracy. The objectives of this study were to 1) develop a multi-modal deep learning (DL) model that integrates time-series behavioral sensor data with single time-point phenotypic data for postpartum disease prediction, and 2) evaluate the effects of representation learning and class-imbalance handling strategies on disease prediction. Data collected from 9,540 Holstein cows across three commercial dairy farms from 21 days before to 2 days after calving were used to predict morbidity within 30 days postpartum, defined as the occurrence of any uterine disease, metabolic disorder, digestive disorder, or other health conditions, with a disease incidence of 27.8%. Behavioral sensor data included daily rumination, activity, days in milk, and temperature-humidity index, and management phenotypes included parity, calving problems, cohort calving number, calf gender, and calving ease score. A multi-modal DL model was developed by fusing time-series features encoded by a time-series DL encoder with single time-point features encoded by a multi-layer perceptron. Three DL encoders were compared, including long short-term memory networks (LSTM), temporal convolutional networks (TCN), and transformer encoders. To assess the effects of representation learning and class-imbalance handling strategies, the multi-modal DL model described above was used as the baseline to implement different model training configurations by adding contrastive multi-modal (CMM) representation learning and one of three class-imbalance handling strategies, including focal loss, positive weighting, and weighted sampling. Model performance was evaluated using both random hold-out cross-validation (CV) with a 70-30 split and cross-farm CV. Across all configurations, the transformer encoders consistently outperformed LSTM and TCN. Under hold-out CV, the transformer encoders with a baseline configuration achieved an area under the precision-recall curve (AUPRC) of 0.693 and an area under the receiver operating characteristic curve (AUROC) of 0.801. Incorporating CMM improved AUPRC and AUROC to 0.719 and 0.808, respectively, and adding focal loss further increased them to 0.731 and 0.826. Under cross-farm CV, the transformer encoders with both CMM and focal loss achieved the highest AUPRC of 0.647 and AUROC of 0.776. These results demonstrate that the proposed multi-modal integration can effectively support postpartum disease prediction. Combining representation learning and class-imbalance handling strategies can further improve minority disease case prediction. We conclude that the proposed approach can improve early disease detection and could potentially be implemented in practical on-farm dairy settings to support timely health management.

Keywords: 2026

How to Cite:

De Castro, A., Chebel, R. & Yu, H., (2026) “Utility of multi-modal integration for predicting postpartum diseases in dairy cattle with imbalanced data”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2285055. doi: https://doi.org/10.31274/wcgalp.23630

Rights: 1