This research aimed: (i) to evaluate on-farm (FARM data) multiparametric models developed under controlled experiment (INRAE data) and based on non-invasive indicators to detect subacute ruminal acidosis (SARA) in dairy cows. We also aimed to recover high discrimination capacity, if needed, by (ii) building new models with combined INRAE and FARM data; and (iii) enriching the models increasing from 2 to 5 indicators per model. For model enrichment, we focused on indicators determinable on-farm by quick and inexpensive routine analysis. Fifteen commercial dairy farms were selected to cover a wide range of SARA risk. In each farm, four Holstein early-lactating healthy primiparous cows were selected based on their last on-farm recording of milk yield and somatic cell count analysis. Cows were equipped with a reticulo-rumen pH sensor. The pH kinetics were analysed over a subsequent 7-day period. Relative pH indicators were used to classify cows with or without SARA. Milk, blood, faeces, and urine were collected for analysis of the indicators included in the models developed by Villot et al. (2020) on INRAE data that were externally evaluated using FARM data. Then, new models based on the same indicators were developed combining INRAE and FARM data to test whether a possible loss in performance was due to a limited validity domain of model by Villot et al (2020). Finally, the models developed combining INRAE and FARM data were adapted to the on-farm application and enriched by increasing indicators from 2 to 5 per model using linear discriminant analysis and leave-one-out cross-validation. The sensitivities (truepositive rate) in external evaluation on FARM data were substantially lower than those from crossvalidation by Villot et al. (2020) (range: 0.1–0.75 vs 0.79–0.96, respectively), and the specificities (true-negative rate) showed a larger range with lower minimum values (range: 0.18–1.0 vs 0.62–0.97, respectively). The sensitivities of new models developed combining INRAE and FARM data ranged from 0.63 to 0.77. Models involving blood cholesterol, b-hydroxybutyrate, haptoglobin, milk and blood urea, and models involving milk fat/protein ratio, dietary starch proportion, and milk fatty acids had the highest performances, whereas models including sieved faecal residues and urine pH had the lowest. Enriching models to three indicators per model improved sensitivity and specificity, but the inclusion of more indicators was less or not effective. Larger field trials are required to validate our results and to increase variability and validity domain of models.