The accuracy of radiological diagnostics is crucial for providing effective medical treatment, ensuring that patients receive timely and appropriate care. Despite the efforts of numerous medical studies to identify radiological errors, this has a negligible effect, and the percentage of annual errors does not decrease. To a large extent, this is explained by the constancy of human perceptual abilities. This study investigates the idea of analyzing longitudinal gaze paths of radiologists and finding consistent patterns associated with diagnostic errors. Three information sources are synergistically studied, namely target medical image, gaze fixation points, and accumulated gaze maps. Methodologically, a transformer was trained on the target image using the gaze map as the attention input. A graph neural network was developed to find the patterns in gaze fixation regions on the image and fixation statistical features. The adjacency of the graph was based on the spatial and temporal differences between fixations. The resulting framework was trained on 3949 chest radiograph readings performed by four radiologists, who on average made some decision errors in 20% of readings. The framework predicted the errors with 0.69 accuracy and 0.74 ROC AUC, which compares favorably to alternative approaches.