| Title |
Machine learning based detection of subacute ruminal acidosis in early lactation dairy cows using multi-sensor behavioral, physiological, and milk production data |
| Authors |
Grigė, Samanta ; Girdauskaitė, Akvilė ; Grigas, Ovidijus ; Japertas, Sigitas ; Malašauskienė, Dovilė ; Televičius, Mindaugas ; Urbutis, Mingaudas ; Džermeikaitė, Karina ; Krištolaitytė, Justina ; Antanaitis, Ramūnas |
| DOI |
10.1038/s41598-026-55162-z |
| Full Text |
|
| Is Part of |
Scientific reports.. Berlin : Nature portfolio. 2026, Early access, p. 1 |
| Keywords [eng] |
Biosensors ; Dairy cattle ; Innovations ; Machine learning |
| Abstract [eng] |
Subacute ruminal acidosis (SARA) is a common metabolic disorder in early lactation dairy cows that negatively affects rumen function, milk production, and animal welfare. Early identification remains challenging because clinical signs are often subtle and transient. The aim of this study was to evaluate whether multi sensor behavioral, physiological, and milk-production data could be used to identify cows experiencing SARA using machine-learning approaches. The study included early-lactation Holstein cows during the first 100 days in milk. The final dataset comprised 636 cow-day observations, including 134 SARA cases and 502 clinically healthy controls. Cow identification numbers were additionally used as grouping variables during cross-validation to ensure complete separation of individual animals between training and validation subsets. SARA was defined based on continuous ruminal pH measurements, where cows were classified as SARA when ruminal pH remained between 5.2 and 5.8 for at least 180 min per day. Sensor derived variables included rumination time, activity, water intake, reticulorumen temperature, milk yield, and milk composition obtained from intraruminal boluses and an in-line milk analyzer. Six supervised machine learning classifiers were developed to classify SARA status based on combined sensor data. Among the evaluated models, SVM demonstrated the highest discriminatory performance, which achieved an area under the curve (AUC) of 0.97, accuracy of 0.95, sensitivity of 0.86, and specificity of 0.98 under repeated cow-level grouped cross-validation. Random forest showed similar performance (AUC = 0.97; accuracy = 0.93 and 0.98, respectively). Across all models, specificity was consistently higher than sensitivity, indicating that healthy cows were classified more accurately than SARA cases. These results demonstrate that integrated behavioral, physiological, and milk production data obtained from automated sensor systems can support classification of cows experiencing SARA under commercial farm conditions. The findings support the potential of multi sensor monitoring systems combined with machine learning classifiers as a tool for automated detection of rumen health disturbances in precision dairy farming. |
| Published |
Berlin : Nature portfolio |
| Type |
Journal article |
| Language |
English |
| Publication date |
2026 |
| CC license |
|
