The objective of this research was to evaluate the efficiency of electronic measurement of activity and lying behavior by ALT-pedometer to recognize different behavior patterns between non-lame and lame cows. The sensors were used to measure the activity and lying behavior, including the total time spent lying down, the number of lying bouts, the duration of each bout for individual cows and maximal/minimal bout duration. A total of 30 lactating Holstein dairy cows were selected based on their locomotion score (NRS ≤ 2). These cows were gait scored according to a 5-point numerical rating system (NRS) and categorized during the experiment as NRS ≤ 2, NRS = 3, NRS = 3.5. This resulted in a dataset of 549 labeled days from eleven cows in total, with approximately the same amount of lame and non-lame days. Huge differences in daily behavior between individual cows were observed. Those differences were significantly larger than the change in daily behavior caused by lameness for each cow. Therefore, it was concluded that thresholds and usage of the absolute values were not feasible to predict lameness for all cows. Hence, instead of using absolute measurements for prediction, the deviation from normal behavior was used for classification. As this deviation was in some features equally likely to differ in positive and negative direction, non-linear prediction models had to be used. In addition, single features were not informative enough to reveal lameness and thus a model combining all features for prediction was necessary. For classification, Support Vector Machines with an RBF-kernel were used. In contrast to a prediction accuracy of 65% from the model derived for absolute values, we were able to predict lameness with an accuracy of 76% using the deviation from normal behavior as features. Our results demonstrate that ALT-pedometer measurements in combination with machine learning tools have the potential to detect lameness accurately on-farm.