Employing Eye Tracking to Assess Proficiency Level in Manufacturing Tasks via Machine Learning

Assessing worker skill is essential for optimizing high-complexity, low-output production tasks that are consequently difficult to automate for increased efficiency over human work. Identifying specific, accessible metrics to identify workers as beginners versus experts may also help design more efficient training regimens, as well as help uncover the determinants of skill. In this study, atemporal (timestamp removed) eye tracking data from 16 subjects performing a soldering task were analyzed using a variety of machine learning models, including k-nearest neighbors (KNNs) and decision trees, to examine if worker task skill could be assessed from nonsequential eye movement and pupil size data alone. We further investigated whether feature extraction via principal component analysis (PCA) could be used to improve the performance, efficiency, and robustness of the prediction models. PCA was selected due to its algorithmic efficiency and previously demonstrated ability to improve the performance of tree algorithms due to the alignment of data on independent axes. We find that fine-tree models trained on 95% PCA data had the most consistent performance classifying expert and beginner sessions across unseen sessions from training subjects and sessions from unseen testing subjects (68.88% mean, 80.97% median session accuracy from training subjects; 72.64% mean, 75.13% median session accuracy from testing). This model also showed strong robustness to outlier trends, suggesting it was able to extract generalizable eye tracking trends that do not rely on sequential context but are still indicative of worker skill.