As eye-tracking technology gains prominent attention in Information Visualizations (InfoVis) research, the need for high accuracy and precision in eye-tracking data becomes increasingly critical. Gaze estimation error is a concept used in eye-tracking, that refers to the difference between the estimated and the true gaze position and direction. Unlike interactions involving physical contact, the visual focus can only be determined through estimation, making gaze estimation errors inevitable. To minimize these errors, calibration is typically performed, where users are asked to look at five or more points on the screen to establish baseline data for ground truth. All following eye-tracking experiments are then based on this calibration data. However, there are situations where calibration is not practical, such as in some remote or online studies or during dynamic activities, where the calibration process cannot be reliably controlled. To address this limitation, this paper proposes VisCaiNet, a deep-learning model that predicts gaze estimation error through post-hoc analysis, using duration, scanpath, scanpath length, and Hit Any AOI Rate (HAAR) as input features, with calibration error as the output. It can effectively discern between high and low-quality gaze data based on the predefined calibration criteria, offering a solution to the challenges posed by variable conditions and unfeasible calibration situations.