367. Multi-modal Machine Learning for the Early Detection of Subclinical Ketosis in Dairy Cattle.

Abstract Subclinical ketosis (SCK) is a prevalent metabolic disorder in transition dairy cows causing significant economic and welfare impacts, yet early detection remains challenging due to absent clinical signs and disease complexity. Two complementary projects developed multi-modal machine learning systems for prepartum prediction of postpartum SCK risk. The first integrated depth image-derived body shape features (extracted via CNN, anatomical keypoint sampling, and CNN-RNN), wearable sensor behavioral data, cow history, and text embeddings from farm management notes using generic and fine-tuned LLMs; Random Forest models combining image and tabular data achieved F₁ = 0.706, with fine-tuned LLM text embeddings further improving sensor-based models (F₁ 0.681 vs. 0.655). The second project added genomic data (78,964 SNPs reduced to 128 dimensions via UMAP) and introduced a cloud-based modular pipeline for automated image processing and data integration; among three fusion strategies tested, late fusion achieved the highest performance (F₁ up to 0.750) for binary SCK classification. Together, these projects show that integrating phenomic and genomic data through modern machine learning and fusion techniques substantially improves early SCK detection, supporting preventive health management in dairy farms.