Evaluación de la separabilidad en datos obtenidos de jóvenes durante la Iowa Gambling Task a través de algoritmos de agrupamiento

Abstract

The Iowa Gambling Task (IGT) is a neuropsychological test designed to assess decision-making under uncertainty. Participants select cards from four decks, two yielding sustained gains and two leading to cumulative losses, allowing the observation of strategy learning over time. The IGT has become a key instrument for studying how individuals face scenarios involving risk and reward. Traditionally, it has been used to predict behavioral performance (mainly balance and net score) and to compare neurotypical populations with clinical groups. However, some studies have shown considerable variability even among neurotypical participants, suggesting that behavioral metrics alone are insufficient to explain differences in decision-making strategies. Incorporating EEG recordings with unsupervised learning techniques provides an opportunity to uncover latent patterns and gain a more comprehensive perspective on this heterogeneity. This study evaluated the separability of behavioral, sociodemographic, and EEG data from neurotypical young adults performing the IGT, by implementing and comparing clustering algorithms. A database integrating these data was analyzed through cluster tendency tests, followed by the application of K-Means, Density-Based Spatial Clustering of Applications with Noise (DBSCAN), Gaussian Mixture Models (GMM), and Agglomerative Hierarchical Clustering. Cluster quality was assessed with the Silhouette coefficient and Rand Index to examine both internal cohesion and correspondence with external variables. The Silhouette coefficient showed positive values across all algorithms (0.34–0.77), with Blocks 3 and 5 achieving the highest scores (≥0.64 in K-Means, GMM, and hierarchical clustering), coinciding with the stage where participants began to identify advantageous decks. Block 4 showed the lowest values, while DBSCAN exhibited limited performance. Agglomerative hierarchical clustering was the most consistent and achieved the best overall results. The Rand Index revealed moderate agreement with sociodemographic variables (0.49–0.52), indicating these did not clearly determine data structure. Greater agreement with behavioral variables was observed in early blocks, especially balance in Block 3, suggesting a key moment in adopting more defined strategies. However, this trend was not sustained later, while net scores declined after the first block. Overall, the findings confirm that unsupervised learning algorithms are valuable tools for analyzing neurocognitive data, as they reveal differences in decision-making strategies not explained solely by behavioral or sociodemographic variables. This replicable approach represents a significant contribution to the multidisciplinary study of decision-making.