BACKGROUND: Early prediction of progression in dementia is of major importance for providing patients with adequate clinical care, with considerable impact on the organization of the whole healthcare system. AIMS: The main task is tailoring robust and consolidated machine learning models to detect which neuropsychological tests are more effective in predicting a patient's mental status. In a translational medicine perspective, such identification tool should find its place in the clinician's toolbox as a support throughout his daily diagnostic routine. A second objective involves predicting the patient's diagnosis based on the results of the cognitive assessment. METHODS: 281 patients with MCI or dementia diagnosis were assessed through 14 commonly administered neuropsychological tests designed to evaluate different cognitive domains. A suite of machine learning models, trained on different subsets of data, was used to detect the most informative tests and to predict the patient's diagnosis. Two external validation datasets containing MMSE and FAB tests were involved in this second task. RESULTS: The tests qualitatively and statistically associated to a cognitive decline are MMSE, FAB, BSTR, AM, and VSF, of which at least three were considered the most informative also by machine learning. 73% average accuracy was obtained in the diagnosis prediction on three subsets of original and external data. DISCUSSION: Detecting the most informative tests could reduce the visits' time and prevent the cognitive assessment from being biased by external factors. Machine learning models' prediction represents a useful baseline for the clinician's actual diagnosis and a reliable insight into the future development of the patient's cognitive status.