Shafiei Somayeh B, Shadpour Saeed, Mohler James L
Roswell Park Comprehensive Cancer Center, USA.
University of Guelph, Canada.
Hum Factors. 2025 May;67(5):464-484. doi: 10.1177/00187208241285513. Epub 2024 Sep 26.
ObjectiveWe aimed to develop advanced machine learning models using electroencephalogram (EEG) and eye-tracking data to predict the mental workload associated with engaging in various surgical tasks.BackgroundTraditional methods of evaluating mental workload often involve self-report scales, which are subject to individual biases. Due to the multidimensional nature of mental workload, there is a pressing need to identify factors that contribute to mental workload across different surgical tasks.MethodEEG and eye-tracking data from 26 participants performing Matchboard and Ring Walk tasks from the da Vinci simulator and the pattern cut and suturing tasks from the Fundamentals of Laparoscopic Surgery (FLS) program were used to develop an eXtreme Gradient Boosting (XGBoost) model for mental workload evaluation.ResultsThe developed XGBoost models demonstrated strong predictive performance with values of 0.82, 0.81, 0.82, and 0.83 for the Matchboard, Ring Walk, pattern cut, and suturing tasks, respectively. Key features for predicting mental workload included task average pupil diameter, complexity level, average functional connectivity strength at the temporal lobe, and the total trajectory length of the nondominant eye's pupil. Integrating features from both EEG and eye-tracking data significantly enhanced the performance of mental workload evaluation models, as evidenced by repeated-measures t-tests yielding -values less than 0.05. However, this enhancement was not observed in the Pattern Cut task (repeated-measures t-tests; > 0.05).ConclusionThe findings underscore the potential for machine learning and multidimensional feature integration to predict mental workload and thereby improve task design and surgical training.ApplicationThe advanced mental workload prediction models could serve as instrumental tools to enhance our understanding of surgeons' cognitive demands and significantly improve the effectiveness of surgical training programs.
目的
我们旨在利用脑电图（EEG）和眼动追踪数据开发先进的机器学习模型，以预测与进行各种手术任务相关的心理负荷。
背景
评估心理负荷的传统方法通常涉及自我报告量表，这些量表容易受到个体偏差的影响。由于心理负荷的多维度性质，迫切需要确定在不同手术任务中导致心理负荷的因素。
方法
来自26名参与者的EEG和眼动追踪数据被用于开发一个用于心理负荷评估的极端梯度提升（XGBoost）模型。这些参与者执行了达芬奇模拟器中的拼板和环形行走任务，以及腹腔镜手术基础（FLS）程序中的图案切割和缝合任务。
结果
所开发的XGBoost模型表现出强大的预测性能，拼板、环形行走、图案切割和缝合任务的预测值分别为0.82、0.81、0.82和0.83。预测心理负荷的关键特征包括任务平均瞳孔直径、复杂程度、颞叶的平均功能连接强度以及非优势眼瞳孔的总轨迹长度。重复测量t检验得出p值小于0.05，这证明整合EEG和眼动追踪数据的特征显著提高了心理负荷评估模型的性能。然而，在图案切割任务中未观察到这种增强（重复测量t检验；p>0.05）。
结论
这些发现强调了机器学习和多维度特征整合在预测心理负荷方面的潜力，从而改善任务设计和手术培训。
应用
先进的心理负荷预测模型可以作为有用的工具，增强我们对外科医生认知需求的理解，并显著提高手术培训项目的有效性。
