Lameness is a very common problem in breeding sows, which often goes undetected for long periods of time. This can have severe consequences for animal welfare and has impact on the productive performance of sows. Automatic lameness detection could help pig farmers to recognize and treat the problem sooner. The SowSIS consists of 4 force plates, built into an electronic sow feeder, providing non-invasive ‘stance’-output for each leg. Stance information variables can be extracted from the data. In a previous study stance variables that could discriminate between 4 lame and non-lame sows were identified. A pilot study was performed with automatic data-collection for 3 groups with a total of 53 gestating group-housed sows. All sows’ gait were visually scored twice a week using a 150mm tagged visual analogue scale to determine their lameness status. This gait score (GS) was used as a reference and compared to the SowSIS measurements. Using Multilevel Linear Regression (MLR), with sow as a random factor to correct for repeated measurements, variables were first tested univariable to identify those with significant influence on GS. Subsequently, these variables were tested using multivariable MLR to determine which variables to use in the prediction model. The variables used were: mean relative weight (RW) on the left, minimum RW on the left, mean of all leg weight ratios (RW of lightest leg/heaviest leg per pair of legs) and the kick frequency of all legs (kicks/min). With the model estimates, it was determined whether the model would classify a sow as lame or not-lame using GS ≥ 55 as the cut-off value. Performance of the model was 69% sensitivity, 86% specificity, 67% lame predictive value and 87% not-lame predictive value. Further analysis to refine interpretation of the model is currently ongoing. The final results will be presented at the conference.