Predictive models of lameness in dairy cows achieve high sensitivity and specificity with force measurements in three dimensions

Abstract Text:

Lameness remains a significant cause of production losses and a growing welfare concern across the dairy industry. Metabolic, nutritional, and housing factors interact to sustain a steady increase in the prevalence of lameness driving a growing need for automated and continuous methods of lameness detection. A force-plate system restricted to the vertical (z) dimension yielded a high specificity, but low sensitivity of detection. The objective of this study was to determine the effect of supplementing the vertical dimension with the transverse (x) and longitudinal (y) dimensions on detection accuracy. We employed a parallel, force-plate system to measure the ground reaction forces (GRFs) across three orthogonal directions (3D). GRFs for randomly selected cows (n=83) were recorded and a clinical diagnosis of lameness was generated using locomotion score, lesion diagnosis, lesion score, and claw and interdigital integument pain score. Logistic regression was employed to characterize the relationship between the clinical characteristics and the GRFs across all 3 orthogonal dimensions to generate a statistical algorithm for the probability of lameness. Misclassification error was estimated using a modification of the Leave-One-Out (LOO) method of cross-validation. LOO cross-validation trains the model using all but a single run. We modified LOO to leave out all runs except for those from a single cow, Leave-One-Cow-Out (LOCO), to use as the training data and tested the resulting model using the runs of the cow not used in model development.

This preliminary study determined that 76 variables across all 3 dimensions resulted in a model with 90% sensitivity, 93% specificity, and 98% area under the receiver operating curve (AUC). Furthermore, all three dimensions were both necessary and sufficient to accurately establish the probability of lameness (Table 1).

Table 1. Model performance using various combinations of measurement directions (1-degree, 4-knot spline transformation). Results have been ordered by increasing AUC.

Measurement Direction (including Stance Time)	TN	FP	FN	TP	Sensitivity	Specificity	AUC
x	213	44	94	45	0.32	0.83	0.59
z	212	45	87	52	0.37	0.82	0.62
x, z	210	47	61	78	0.56	0.82	0.73
y	218	39	77	62	0.45	0.85	0.75
y, z	222	35	53	86	0.62	0.86	0.79
x, y	221	36	50	89	0.64	0.86	0.83
x, y, z	239	18	14	125	0.90	0.93	0.98

x = transverse (medial-lateral) direction

y = longitudinal (cranial-caudal) direction

z = vertical (weight) direction

Keywords:

lameness, animal welfare, 3-dimensional ground reaction forces