TL;DR
This study investigates how different causes of drowsiness (like fatigue or sleep deprivation) and assessment methods (subjective vs. objective) affect physiological signals used for monitoring. The findings reveal that distinct drowsiness inducers lead to varied physiological responses, and objective assessments are more sensitive. This work is crucial for developing more robust and generalizable drowsiness detection systems for safer driving.
What Problem Does This Paper Address?
Drowsiness while driving is a major safety concern, leading to numerous accidents, injuries, and fatalities. Physiological sensors offer a privacy-preserving alternative to camera-based monitoring for detecting drowsiness. However, existing drowsiness detection models often struggle with generalizability because studies use different methods to induce drowsiness (e.g., sleep deprivation, monotonous driving) and assess its severity (e.g., self-reports, behavioral cues). This inconsistency means that models trained on one dataset might not perform well on data collected under different conditions, hindering the development of reliable, real-world drowsiness monitoring systems. This paper aims to bridge this gap by analyzing how different drowsiness inducers and assessment methods influence physiological signals, seeking to identify robust markers for generalizable drowsiness detection.
Methods
The researchers analyzed data from four distinct datasets, each employing different drowsiness induction methods and assessment criteria. They focused on physiological signals from electrocardiography (ECG), electrodermal activity (EDA), and respiration (RESP).
The core methodological steps included:
- Data Preprocessing: Standardizing all signals to a 100 Hz sampling frequency, applying low-pass and band-pass filters to remove noise from ECG, EDA, and RESP signals.
- Feature Extraction: Extracting 32 physiological features from the preprocessed signals.
- Drowsiness Labeling: For datasets using subjective measures like the Karolinska Sleepiness Scale (KSS), they binarized sleepiness into "drowsy" (KSS ≥ 7) and "awake" (KSS < 6) categories.
- Addressing Multicollinearity: Calculating Pearson correlations between features and either aggregating or discarding highly correlated variables (r > 0.8 or r < -0.8, p < .05) to select 14 independent features.
- Model Building: Employing binary logistic regression models (using the GENMOD procedure with generalized estimation equations to account for repeated measures) to identify physiological metrics associated with drowsiness. Five models were built: four for individual datasets and one for an aggregated dataset. A backward stepwise selection based on QIC was used to obtain the best-fitted models.
Datasets
The study utilized four datasets:
- Fatigueset (2021): Focused on physical fatigue induced by exercise. Drowsiness was assessed objectively.
- AdVitam (2023): Involved sleep-deprived drivers in a Level 3 automated driving scenario. Drowsiness was assessed subjectively using the KSS.
- MCDD (2024): Included drivers performing cognitive tasks during Level 3 automated driving. Drowsiness was assessed subjectively using the KSS.
- Ayas et al. (2024) (Private Dataset): Involved drivers in a monotonous Level 2 automated driving scenario. Drowsiness was assessed objectively by independent raters based on facial-video indicators.
Key Insights and Ideas
- Inducer-Specific Responses: Different ways of inducing drowsiness (physical fatigue, low arousal, sleep deprivation, mental fatigue) lead to distinct physiological responses. This highlights why a model trained on one type of drowsiness might not generalize well to others.
- Objective vs. Subjective Assessment: Objective assessments (like rater annotations) were significantly more sensitive in detecting drowsiness than subjective self-reports. This suggests objective measures can capture subtle physiological changes missed by self-assessment.
- Robust Physiological Markers: Despite variations, certain physiological features consistently emerged as reliable indicators of drowsiness across different datasets and conditions. These include higher heart rate, more stable cardiac activity, reduced respiratory amplitude, and lower skin electrical activation.
- Mental Fatigue is Tricky: Mental fatigue was less likely to be detected as drowsiness compared to other inducers, suggesting it might present unique physiological signatures that are harder to capture with current methods.
Key Results
- Physical Fatigue & Low Arousal: Both were linked to changes in RESP-RR and EDA-Min tonic, though in opposite directions.
- Sleep Deprivation & Mental Fatigue: Both were associated with ECG and EDA indicators. RESP-RMSSD was linked to sleep deprivation, but no RESP feature was linked to mental fatigue.
- Objective Assessment Advantage: Objective assessments were significantly more likely to identify participants as drowsy (OR = 12.16).
- Drowsiness Inducer Impact: Mental fatigue had a lower likelihood of being identified as drowsy compared to physical fatigue (OR = 0.04), sleep deprivation (OR = 0.19), and low arousal (OR = 0.04). Conversely, low arousal (OR = 5.23) and physical fatigue (OR = 4.65) drivers were more likely to be identified as drowsy compared to sleep deprivation.
- Consistent Markers: Higher heart rate, more stable cardiac activity, reduced respiratory amplitude, and lower skin electrical activation were identified as robust markers across various conditions.
Limitations and Future Work
Limitations:
- Dataset Reliance: The study's findings are limited by the existing datasets, which may not fully represent the diversity of real-world driving scenarios and participant populations.
- Simulated Environment: All datasets were collected in simulated environments. This raises questions about the generalizability of the findings to naturalistic driving conditions.
- Physical Fatigue Dataset: The lack of a public dataset specifically for drowsy driving induced by physical fatigue meant the researchers had to use a non-driving dataset for this condition, which might limit direct comparisons.
Future Work:
- Naturalistic Driving Studies: Future research should focus on validating these physiological features in naturalistic driving environments.
- Diverse Datasets: The authors call for future drowsiness detection research to incorporate diverse drowsiness inducers and assessment methods to improve model generalizability.
- Balanced Assessment: Future models and assessment protocols should aim to balance the sensitivity and specificity of both subjective and objective metrics to mitigate miss and false alarm rates.
- Model Complexity: Focusing on the identified robust indicators could help streamline model complexity and enhance external validity for deployment across various driving scenarios.
Summary updated