An automated method for determining whether dairy cows with subclinical mammary infections recover after antibiotic treatment would be a useful tool in dairy production. For that purpose, inline _L-lactate dehydrogenase (LDH) measurements was modeled using a dynamic linear model; the variance parameters were estimated using the expectation-maximization algorithm. The method used to classify cows as infected or uninfected was based on a multiprocess Kalman filter. Two learning data sets were created: infected and uninfected. The infected data set consisted of records from 48 cows with subclinical Staphylococcus aureus infection from 4 herds collected in 2010. The uninfected data set came from 35 uninfected cows collected during 2013 from 2 herds. Bacteriological culturing was used as gold standard. To test the model, we collected data from the 48 infected cows 50 d after antibiotic treatment. As a result of the treatment, this test data set consisted of 25 cows that still had a subclinical infection and 23 cows that were recovered. Model sensitivity was 36.0% and specificity was 82.6%. To a large extent, _L-lactate dehydrogenase reflected the cow's immune response to the presence of pathogens in the udder. However, cows that were classified correctly before treatment had a better chance of correct classification after treatment. This indicated a variation between cows in immune response to subclinical mammary infection that may complicate the detection of subclinically infected cows and determination of recovery.