This study aimed to explore the performances of Facial Action Units (FAUs) to detect pain in cows under lipopolysaccharide (LPS) challenge though a six-step process based on expression, selectivity (construct validity, responsiveness), intra-and inter-observer reliability, and measurement error (specificity, Sp; sensitivity, Se). Twenty-seven cows received an intra-mammary infusion of 25 & mu;g E. coli LPS in one healthy udder quarter. Then, 14 cows received a placebo (LPS cows) and 13 cows received 3 mg/kg BW ketoprofen i.m. (LPS+NSAID cows). Each cow's face was video-recorded for 40 s at three time points before (T-21 = 11:30 h, T-19 = 13:30 h, T-16 = 16:30 h) and after (T3 = 11:30 h, T5 = 13:30 h, T8 = 16:30 h) infusion. Three trained observers scored the duration, frequency or presence/absence of 43 FAUs on 40 second video segments. We kept only the selective FAUs (construct validity) and transformed them into binary variables (below / above a threshold determined by ROC curves and Youden index). Intra-and inter-observer reliability were assessed by percentage of agreement (PA) and Fleiss' kappa (k). We calculated Sp and Se. The process therefore consisted in a waterfall method with expression, selectivity, intra-observer reliability being an eliminative step, while inter-observer reliability, Sp and Se were not. FAUs were kept if expressed (>5% of the videos), and kept for intra-observer reliability if PA & GE; 75% and k & GE; 0.41. Two too rarely expressed FAUs were excluded. LPS infusion induced changes in 7 FAUs of orbital, auricular, and mouth-and-muzzle regions. Compared to before challenge (T-19), at T5, LPS cows spent signifi-cantly less time with 'muzzle in motion' (P = 0.045), tended to display more 'nostril dilation' (P = 0.097), spend more time with 'motionless muzzle' (P = 0.068) and less time in 'ear: position 8 & PRIME; (i.e. backwards /central/pinna to the side) (P = 0.057). At T8, LPS cows spent significantly less time with 'eye open' (P = 0.036) and tended to less frequently display 'eye blinking' (P = 0.071) and 'eye movements' (P = 0.091) compared to T-16. Four of these 7 FAUs ('eye open', 'motionless muzzle', 'muzzle in motion', 'nostril dilatation') satisfied all following steps of the process except sensitivity. Two other FAUs ('eye blinking', 'eye movements') satisfied intra-and inter -observer reliability; depending on the time point considered they were either sensitive or specific but not both simultaneously. The last FAU ('ear: position 8 & PRIME;) satisfied intra-observer reliability and sensitivity but not inter-observer reliability nor specificity. This study identified 7 FAUs as potential candidate for detecting mild pain associated with induced inflammatory mastitis in dairy cows.