To reach a broader population during Mild Cognitive Impairment population screenings, there is a need for neuropsychological test batteries that are quick to administer and evaluate. To achieve that, all test scores must be informative and non-redundant. This study used machine learning methods to analyze the test scores in a neuropsychological database in terms of relevance and collinearity. We examined a database of 520 assessments with diagnoses of either healthy or MCI, evaluated using a neuropsychological test battery consisting of fourteen tests, from which we will use twenty-three test scores. First, we assessed both the correlation between test scores and the diagnostic dependence of each score. This led us to define four groupings of test scores (very high relevance, very high and high relevance, all except low-relevance, and all test scores) that exclude redundant test scores within each group, and a fifth group that includes all test scores, even redundant ones, as a baseline. Those test score groups were analyzed using well-established machine learning methods. Based on the machine learning analysis, the fifth group had the best performance, likely because of the redundant test scores. Among the groupings that exclude redundant test scores, we found that the most promising group was the one with all test scores except the low-relevant ones, though it was almost tied in performance with the group containing all test scores.