Interpretable Driver Fatigue Detection Using SHAP Analysis of Multiple Entropy EEG Features

Hu, Jianfeng; Gao, Jing · 2026 · Crossref

DOI: 10.21203/rs.3.rs-10209165/v1

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the lack of clinical interpretability in EEG-based driver fatigue detection models, which often function as "black boxes" despite achieving moderate classification accuracy. While entropy-based features capture signal complexity changes associated with fatigue, existing methods fail to provide transparent explanations for classification decisions, hindering trust and regulatory acceptance in safety-critical applications. The authors aim to identify reliable, interpretable biomarkers by applying SHapley Additive exPlanations (SHAP) to tree-based classifiers trained on multi-entropy EEG features. The researchers conducted a simulated driving experiment with 26 healthy subjects (aged 17–24) to induce fatigue through sustained attention tasks. EEG signals were recorded from 30 channels and segmented into 1-second epochs. Six entropy measures—Approximate, Sample, Fuzzy, Permutation, Spectral, and Kolmogorov—were computed for each epoch, resulting in 180 features per second. The study employed Leave-One-Subject-Out (LOSO) cross-validation to assess cross-subject generalization using Random Forest, XGBoost, and LightGBM classifiers, with SVM and LDA serving as baselines. TreeSHAP was used to quantify feature contributions, and cross-subject consistency was validated via Bonferroni-corrected permutation testing. Results indicated that LightGBM achieved the highest mean AUC (0.678), though performance across tree-based models was statistically similar, suggesting inherent cross-subject variability limits accuracy. SHAP analysis identified Fuzzy Entropy (FuEn) as the most informative entropy type, with FuEn at the Pz channel (FuEn_Pz) emerging as the top global feature (mean |SHAP| = 0.02013). FuEn_Pz demonstrated strong cross-subject consistency, ranking in the top three for 65.4% of subjects (corrected p < 0.001). The model exhibited a 28.9% false-negative rate for fatigue detection, indicating a significant safety gap. SHAP rankings diverged from univariate effect sizes, revealing that FuEn_Pz’s importance stemmed from non-linear interactions within the multivariate feature space rather than isolated discriminative power. The study concludes that SHAP enables the extraction of dependable, clinically meaningful biomarkers from pre-computed EEG features, with Fuzzy Entropy at centro-parietal channels proving most robust. The findings highlight that while entropy features offer interpretability, the current false-negative rate necessitates further refinement, such as ensemble methods or multimodal fusion, before clinical deployment. The work underscores the value of explainable AI in bridging the gap between machine learning performance and clinical trust in physiological monitoring systems.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• drowsiness detection algorithms

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: physiological data