Bayesian Network Structures for Early Diagnosis of MCI Using Semantic Fluency Tests

The early detection of MCI has become one of the main focuses in research since it allows early treatment to improve patient quality of life. Currently, several studies seek to combine tests with machine learning systems to increase the efficiency of diagnostics. Bayesian Networks (BN) have the advantage over other machine learning systems of being explainable. However, there is little information on how different network structures affect BN performance in the clinical setting. In this study, we used semantic fluency tests, one of the most widely used neuropsychological tests, to compare the performance of three models: Naive Bayes, BN with the variables grouped by semantic category ( \(\text {BN}_{\text {Test}}\) ), and BN with the variables grouped by type of variables ( \(\text {BN}_{\text {Var}}\) ). The models were compared on three decision thresholds: the standard 0.5, the one that maximizes the model performance, and the one that optimizes the false negatives. The results show that the best conformation of the network corresponds to ( \(\text {BN}_{\text {Var}}\) ) in the three thresholds selected, while ( \(\text {BN}_{\text {Test}}\) ) scores the worst results. This study shows that a proper structure improves the results of the BN concerning the simpler model of Naive Bayes.