Cross-farm generalisation of machine and deep learning models for lameness prediction in multiparous dairy cows

Lameness in dairy cows is a major health and welfare concern and there is interest in the use of pre-recorded data for lameness risk assessment, since on-farm data recording is increasing substantially. The aim of this study was to explore the extent of cross-farm generalisability of models constructed using pre-recorded data, by comparing cross-farm lameness predictions. Data were collated from a group of co-developer farmers participating in ongoing research with Quality Milk Management Services Ltd. The final dataset consisted of 40495 mobility scores and records of farmer recorded lameness events from 2690 animals in 10 herds and information on calving dates and milk production. Three machine learning models, and two deep learning models were tested for ability to determine a mobility score of ≥ 2 (indicating a lame animal) compared with < 2 (indicating a sound animal), using past lameness history, age, parity and past milk production as predictor variables. Across farms, all models were better at predicting non-lame animals than lame animals. Model calibration was better at higher probabilities of lameness, post-hoc exploratory tuning of threshold parameters at farm-level increased the mean sensitivities of the models across farms to 0.64-0.71 from 0.21 to 0.28. Previous lameness history, age and history of sole ulcer (SU) and other lesions were the most consistent predictors of lameness across all model types and farms. Considerable variation in model performance existed between farms; indicting future work needs to determine which farm-specific factors should be recorded to build models for lameness prediction that scale effectively in real-world settings.