Abstract
Hypocalcemia commonly affects dairy cows around calving or at the onset of lactation. An individual risk prediction would allow customized prophylactic measures and targeted intervention. This study aimed to measure concentrations of total (tCa) and ionized calcium (iCa) during the peripartal period, record rumen motility and temperature patterns, and assess the associations between patterns before calving and Ca concentration at calving. A total of 89 calvings from 47 cows and 22 heifers were monitored using reticuloruminal bolus sensors over 60 days prior to the expected calving and 60 days postpartum. Cows with low tCa (<1.8 mmol/L) or iCa (<0.8 mmol/L) a few days before parturition had reduced rumination times compared to cows with normal tCa levels (>2.2 mmol/L). Using a rumination time (RT) threshold of 480 min/day on day -1, hypocalcemia was predicted with 66.7% sensitivity and 96.8% specificity. Additionally, we evaluated a deep learning model trained on external data, which incorporated rumen cycles, reticular contraction duration, and reticular temperature. Despite not being trained on this dataset, the model surpassed the RT thresholding approach, achieving 83.2% sensitivity and 98.2% specificity for tCa-based classification. These results indicate superior performance and greater generalizability of the deep learning approach, highlighting the potential of multi-metric sensor analytics to improve early hypocalcemia risk detection.