驾驶疲劳EEG检测中二分类与三分类的对比研究

Comparative Study of Binary and Ternary Classification in EEG-Based Driving Fatigue Detection

摘要: 针对驾驶疲劳检测中分类策略选择不明确的问题，基于公开的多模态驾驶疲劳表型数据集MPD-DF，采用EEGConformer模型对比二分类(清醒vs.疲劳)与三分类(清醒vs.轻度疲劳vs.中重度疲劳)策略。实验采用样本级分层划分，经数据清洗、标签映射与被试内归一化后，二分类纳入33名有效被试，三分类纳入14名有效被试。结果表明，三分类任务测试集准确率达92.41%，精确率89.82%，召回率93.62%，F1分数91.55%，各项核心指标均优于二分类(准确率90.54%，F1 85.13%)；其中三分类对中重度疲劳的召回率高达97.88%，临床预警价值突出。三分类策略通过保留疲劳程度的层级划分，有效降低了误报率并实现了对疲劳渐进过程的分级刻画，可为驾驶疲劳的分级预警与差异化干预提供可靠的识别依据。

Abstract: To address the ambiguity in classification strategy selection for driving fatigue detection, this study compares binary classification (wakefulness vs. fatigue) and ternary classification (wakefulness vs. light fatigue vs. moderate-severe fatigue) using the EEGConformer model on the public MPD-DF dataset. After data cleaning, label mapping, and subject-wise normalization, the binary task included 33 valid subjects and the ternary task included 14 valid subjects under stratified sample-level partitioning. Results show that the ternary task achieves a test accuracy of 92.41%, precision of 89.82%, recall of 93.62%, and F1-score of 91.55%, all superior to the binary task (accuracy 90.54%, F1 85.13%). Notably, the recall for moderate-severe fatigue in the ternary task reaches 97.88%, demonstrating prominent clinical early-warning value. By preserving the hierarchical division of fatigue levels, the ternary strategy effectively reduces false alarms and enables graded characterization of the fatigue progression process, providing a reliable basis for graded early warning and differentiated intervention in driving fatigue monitoring.