In epilepsy patients, language lateralisation is an important part of the presurgical diagnostic process. Using task-based fMRI, language lateralisation can be determined by visual inspection of activity patterns or by quantifying the difference in left- and right-hemisphere activity using variations of a basic formula [(L-R)/(L+R)]. However, the values of this laterality index (LI) depend on the choice of activity thresholds and regions of interest. The diagnostic utility of the U also depends on how its continuous values are translated into categorical decisions about a patient's language lateralisation. Here, we analysed fMRI data from 712 epilepsy patients who performed a verbal fluency task. Each fMRI data set was evaluated by a trained human rater as depicting left-sided, right-sided, or bilateral lateralisation or as being inconclusive. We used data-driven methods to define the activity thresholds and regions of interest used for LI computation and to define a classification scheme that allowed us to translate the U values into categorical decisions. By deconstructing the LI into measures of laterality (L-R) and strength (L+R), we also modelled the relationship between activation strength and conclusiveness of a data set. In a held-out data set, predictions reached 91% correct when using only conclusive data and 82% when inconclusive data were included. Although only trained on human evaluations of fMRIs, the approach generalised to the prediction of language Wada test results, allowing for significant above-chance accuracies. Compared against different existing methods of U-computation, our approach improved the identification and exclusion of inconclusive cases and ensured that decisions for the remaining data could be made with consistently high accuracies. We discuss how this approach can support clinicians in assessing fMRI data on a single-case level, deciding whether lateralisation can be determined with sufficient certainty or whether additional information is needed.