TL;DR
This paper presents a machine learning approach to detect driver drowsiness using only skin conductance (SC) signals from a wearable wrist device. By developing an improved artifact removal technique and employing ensemble algorithms, the study achieves a promising 89.4% accuracy, demonstrating the potential for comfortable, non-invasive drowsiness detection.
What Problem Does This Paper Address?
Drowsy driving is a significant contributor to road accidents worldwide, leading to millions of deaths and injuries annually. While automatic driver drowsiness detection systems are becoming mandatory in vehicles, many existing solutions rely on intrusive sensors or complex cabin-mounted cameras. This paper addresses the challenge of developing a comfortable, non-invasive, and effective drowsiness detection system by focusing on a single, easily wearable physiological signal: skin conductance (SC). The goal is to leverage SC data from a wrist-worn device to accurately identify when a driver is becoming drowsy, enabling timely warnings to prevent accidents.
Methods
The core of this research lies in its machine learning approach to analyze skin conductance (SC) signals. The authors employed an Empatica E4 wearable device to collect SC data, emulating a realistic scenario where drivers could wear such a device. A key methodological contribution is the development of an improved artifact removal algorithm. This new filter, based on a modified Stationary Wavelet Transform (SWT) with an overlapping window and focusing on the central part of the signal segment, effectively addresses distortions caused by motion artifacts, which are a common problem with wrist-worn sensors.
Following artifact removal, features were extracted from the SC signal, including its components (SCL and SCR). To handle potential redundancy, a supervised feature selection method using Pearson's correlation was applied, reducing the initial 23 features to 18. Three machine learning algorithms—Random Forest, Bagging, and Boosting—were then tested for their ability to classify drowsiness levels into three classes: alert, slightly drowsy, and drowsy. The models were evaluated using 10-fold cross-validation.
A crucial aspect of their work was also comparing the performance of the wearable E4 device with a more invasive, high-fidelity "gold standard" device (ProComp Infiniti) to understand the limitations of wearable sensors. To further enhance the E4's performance, they also explored oversampling the E4's SC signal from its native 4 Hz to 256 Hz, matching the sampling rate of the ProComp device.
Datasets
The study utilized data collected from nine healthy volunteers (four males, five females) aged 20-60, who participated in simulated driving sessions. A smaller dataset of two volunteers was also used for a direct comparison between the wearable E4 and the ProComp Infinity device. The perceived level of drowsiness was recorded using the Karolinska Sleepiness Scale (KSS) every 10 minutes. No new benchmark datasets were introduced; rather, existing data collection protocols were adapted for the study.
Key Insights and Ideas
- The Power of a Single Signal: This research highlights that even a single physiological signal like skin conductance, when properly processed, can be a powerful indicator of driver drowsiness. This simplifies system design and enhances user comfort.
- Artifacts are the Bottleneck: The study strongly emphasizes that the quality of SC signals from wearable devices is heavily impacted by motion artifacts. Developing robust artifact removal techniques is paramount for achieving reliable drowsiness detection.
- Wearable vs. Gold Standard Trade-off: While more invasive, high-sampling-rate devices (like ProComp Infinity) offer superior accuracy, the study demonstrates that with advanced signal processing and oversampling, wearable devices can achieve competitive performance, making them more practical for real-world applications.
- Oversampling Boosts Performance: Increasing the sampling frequency of the wearable device's SC signal through oversampling significantly improved classification accuracy, bridging a substantial gap between the wearable and gold-standard devices.
Key Results
- The developed improved artifact removal filter successfully mitigated distortions caused by motion artifacts, preventing misinterpretation of SC signals and preserving feature integrity.
- When comparing the wearable Empatica E4 with the ProComp Infinity, the ProComp Infinity achieved higher accuracy (e.g., 91.7% for Boosting vs. 83.2% for E4).
- After applying the oversampling technique to the E4's SC signal (increasing sampling frequency to 256 Hz), the classification accuracy significantly improved. The Boosting algorithm achieved the highest accuracy of 89.4%, with precision and recall also around 89.5%.
- This oversampled E4 approach showed an average improvement of almost 5 percentage points compared to the original E4 data.
- The study successfully demonstrated that it is feasible to detect driver drowsiness using only SC signals from a single wrist-worn device in a heterogeneous population.
Limitations and Future Work
- Dataset Size: The authors acknowledge that the experimental dataset, while sufficient for initial validation, could be extended for a more robust and statistically significant evaluation of the drowsiness detection capability.
- Population Heterogeneity: While efforts were made to include a range of ages and genders, a larger and more diverse population would provide a more refined representation of driver behavior.
- Future Work: The authors suggest extending the experimental dataset and focusing on real-world driving scenarios to further validate the system. They also aim to develop a real-time warning system based on these findings.
Summary updated