Neurowearable devices play an important role in monitoring the drowsiness and health status of pilots and drivers. While drowsiness monitoring solutions such as camera-based eye tracking, steering wheel trajectory sensors, or electrophysiological recording devices are widely used in vehicle scenarios, they have their limitations, such as eye tracking being easily obscured by sunglasses and other obstacles. At present, neurowearable devices have a promising future, but most of them are inconvenient to use due to wet electrodes and large size, which limits their daily application.

Recently, a research team at the University of California, Berkeley, developed a novel wearable in-ear ExG sensor system for drowsiness monitoring. The wireless dry electrode earbuds integrate dry electrodes, wireless electronic devices, and offline classification algorithms. Using a support vector machine (SVM) classification algorithm, the research team found that the wireless dry electrode earbuds were comparable to existing wet electrode acquisition systems, such as in-ear EEG and scalp EEG systems, in terms of drowsiness classification accuracy. The application of new wireless dry electrode earbuds will lay the foundation for future covert, wireless, long-term brain monitoring. The findings were published in the journal Nature Communications on August 2, 2024.

Fig1. Conceived in-ear ExG wearables. The system can be comfortably worn in the ear canal throughout the day to record nerve signals for drowsiness detection and provide feedback.

1. Design and manufacture of in-ear ExG system electrodes

Earbud design: Neurowearable devices often require a universal earbud and electrode design to accommodate the long-term wear needs of different groups. The researchers designed small, medium, and large earbuds with electrodes positioned close to the ear canal. The medium-sized earbuds include 4 x 60 mm2 electrodes within the ear canal and 2 x 3cm2 electrodes in the cochlear cymbals and cochlear cavity (Figure 2A). The 3D-printed soft earbud body enhances user comfort and allows the electrodes to move independently to accommodate different ear shapes. This modular design demonstrates the capabilities of the earbud manufacturing process (Figure 2b).

Electrode fabrication: Researchers have developed a low-cost, electroless plating process that enables rapid prototyping of arbitrarily shaped electrophysiology sensors. The electrodes are 3D printed with sandblasting to increase the surface roughness, followed by copper, nickel, and gold metal layers (Figure 2c), which are suitable for dry electrode recording. The nickel layer significantly extends the life of the electrode and simplifies the manufacturing process.

Fig2. Earbuds design, manufacturing and assembly process. A: The earbuds consist of 4 in-ear electrodes and 2 out-of-ear electrodes. The 3D-printed earbuds are assembled by inserting the gold-plated earbuds into a soft, flexible skeleton. b: The extraauricular electrodes are pressed against the cochlear cymbals and cochlear cavity, and the in-ear electrodes are at the entrance to the ear canal. C: Schematic diagram of electrode fabrication. i) The electrodes are 3D printed or molded. ii) The bare electrode is sandblasted and cleaned. iii) Electroless copper plating of the electrodes. iv) Electroless nickel plating. v) Electroless gold plating.

2. Characteristics of electroplating process

Material pickling tests and tape tests: The electrode surface has a high physical and chemical durability. Kapton tape tests show strong adhesion to the substrate with no gold, nickel, or copper peeling. In the pickling test, the electrode was stable in a 1 M nitric acid bath with no corrosion and no microcracks (Figure 3a). The final plating surface roughness increases, showing a low electrode-skin impedance (ESI).

Surface Roughness Characteristics: The surface roughness is slightly reduced during the plating step, but the gold surface is ultimately rougher than the planar surface, increasing the electrode surface area, promoting film adhesion, and reducing ESI (Figure 3b).

Chip resistance: Four-point probe testing shows that the addition of metal to each layer stabilizes the chip resistance and enhances the surface conductivity (Figure 3c).

Bioimpedance of multi-user in-ear electrodes: Impedance spectroscopy tests evaluated the skin impedance of in-ear electrodes and showed an average interface impedance of 120 kΩ at 50 Hz. All measurements were performed using LCR tables and the results were fitted to an equivalent circuit model for future analog front-end design references (Figure 3d-e).

Fig3. Plating surface features. a: Light microscope image of the coating table. b: Flat sample measured by profilometer after each plating step. c: Absolute sheet resistance measurement. D: In-ear electrode-skin impedance size, phase, and fit. e: Constant-phase element electrode model for fitting.

3. Lightweight ExG recording system

The ExG signal is recorded via a WANDmini wireless platform fixed by the headband (Figure 4a). WANDmini is based on a custom neural recording circuit with a sampling rate of 1 kSps and can record 64 fully differential channels. It has a unipolar layout and uses a reference electrode for EEG, EOG, and EMG recording. The system transmits data via Bluetooth Low Energy and has low overall power consumption and can operate for about 44 hours. Its high channel count and low noise floor make it suitable for in-ear EEG prototypes.

Fig4. Experimental setup, documentation, and labeling protocols. A: Participants participate in a game experiment and measure their basic reaction time. The headset's WANDmini is fixed in a 3D-printed case. When the participant is playing a game, the contralateral earbud records and transmits the ExG signal to the base station. b: Recorded ExG signals, reaction time, and Likert scale scores were used to generate features and labels for the brain state classifier. Marked drowsiness events are shaded in green.

4. EEG characteristics and user drowsiness detection

Drowsiness study: Nine subjects wore two earplugs, induced drowsiness through simple games, and recorded basic reaction time, Karolinska Sleepiness Scale (KSS) score, and in-ear ExG data. Reaction time and KSS scores were used to generate alertness, drowsiness labels. All data were used for post-processing and machine learning model training (Figure 4B).

Machine learning algorithm classification process: The training process of ExG data includes data post-processing, feature extraction, and model classification training (Fig. 5a). The data is maximized spatial coverage, bandpass filtering, time window segmentation, and artifact data is discarded. Extract temporal and spectral features associated with drowsiness detection, including eye tracking artifacts and standard EEG band activity: δ (0.05-4 Hz), theta (4-8 Hz), α (8-13 Hz), β (13-30 Hz), and γ (30-50 Hz). After that, logistic regression, SVM, and random forests were used for alertness, sleepiness binary classification.

Three cross-validation techniques were used to estimate model performance: user-specific, leave-one trial-out, and leave-one-user-out cross-validation. To account for the imbalance between sleepiness and alertness classes, each model uses a category balancing scheme, with the weights of the alertness periods inversely proportional to their number (Figure 5b).

Fig5. Drowsy classifier training and validation graphs. a: The ExG data is rereferenced, filtered, and de-artifact, followed by feature extraction and model training. b: Cross-validation is performed in a similar manner, with the processed data entered into three classifiers and classified validation.

5. Machine learning classification results

α modulation ratio: α wave (8-12 Hz) is a spontaneous neural signal that reflects a state of relaxation and is an important spectral feature for the classification of drowsiness. Figure 6a shows an example of α waves for an individual.

Classifier accuracy: In user-specific cross-validation, the average classification accuracy ranges from 77.9% to 92.2%. The average accuracy of leave-one trial-out was higher, ranging from 91.4% to 93.2%, which may be due to the increased amount of training data; The average leave-one-user-out accuracy is 88.1% to 93.3%.

10s vs. 50s Time Feature Window: The study found that in user-specific cross-validation, the 10s window resulted in a significant performance loss, and the average performance of the logistic regression classifier increased from 77.9% to 90.8% when the window size was increased to 50s. In the cross-validation of leave-one trial-out and leave-one-user-out, the feature loss in the 10s window is small, which may be due to the increase in the amount of training data.

Classifier architecture comparison: Logistic regression uses a stochastic mean gradient descent solver and L1 regularization. The SVM uses the RBF kernel, and the training model uses up to 400 support vectors, with the regularization parameter C=1. The random forest model uses 100 trees with a maximum depth of 5. All three models can achieve high accuracy, and the logistic regression model has high computational efficiency and low memory requirements, which is suitable for training and testing with small datasets.

Fig6. EEG measurement and classifier performance. a: Spectrogram of α-wave modulation. b: Logistic regression sleepiness event detection for 10s feature window. c: SVM event detection in the 10s feature window. d: Random forest event detection in a 10s feature window. E, F, G: Event detection for the 50s feature window.

This article presents the design and manufacturing process of the in-ear stem electrodes, as well as the composition and test results of a wireless in-ear ExG recording system for drowsiness detection. The new in-ear sensing system is suitable for different age groups and has the ability to be reused over a long period of time. The results of the machine learning classification accuracy show that dry electrode-based brain state classification has great potential, which not only lays the foundation for the large-scale application of user-friendly wireless in-ear ExG brain-computer interfaces, but also is expected to be used in the future to track long-term cognitive changes caused by diseases such as depression, Alzheimer's disease, or stress.

Reference: 

https://www.nature.com/articles/s41467-024-48682-7

It is only used for academic sharing, if there is any infringement, please leave a message and delete the infringement immediately!

This article is from Xinzhi self-media and does not represent the views and positions of Business Xinzhi.If there is any suspicion of infringement, please contact the administrator of the Business News Platform.Contact: system@shangyexinzhi.com