TL;DR
This paper introduces a novel method for continuously tracking daytime sleepiness using only respiratory and accelerometer data from a wearable sensor. By combining class imbalance handling techniques with a Hidden Markov Model (HMM) and supervised classifiers, they achieve a promising F-measure of 0.7059, demonstrating the potential for unobtrusive, long-term sleepiness monitoring in daily life.
What Problem Does This Paper Address?
Daytime sleepiness is a significant issue, contributing to reduced productivity, accidents, and health problems. While its importance is recognized, long-term, quantitative analysis of sleepiness in everyday situations has been challenging due to the time and effort required for continuous tracking. Existing sleepiness detection technologies often focus on driver drowsiness and rely on complex physiological measurements or laboratory settings, making them unsuitable for daily use. This paper aims to bridge this gap by developing a method for predicting subjective sleepiness levels using readily available data from a wearable sensor, enabling unobtrusive, long-term monitoring in natural environments.
Methods
The authors propose a multi-stage approach to predict daytime sleepiness:
- Data Collection: They collected respiration and accelerometer data from 18 participants engaged in naturalistic tasks (typing and watching videos) over multiple sessions. Subjective sleepiness was recorded using the Karolinska Sleepiness Scale (KSS).
- Preprocessing: Respiratory data was filtered using a bandpass Butterworth filter. Both respiratory and accelerometer data were segmented into 60-second windows with 30-second overlaps.
- Feature Extraction: Nine respiration parameters (e.g., inspiratory duration, expiratory duration, pause duration, areas under curves) and body movement features (derived from accelerometer data) were extracted for each window.
- Feature Selection: A two-step process was employed: first, highly correlated features were removed, and then aggregated feature importance from multiple random forests was used to select the top four features. The selected features were the mean of Inspiratory Duration, mean of Expiratory Duration, minimum of Duration Ratio, and minimum of body-movement parameter.
- Class Imbalance Handling: Recognizing that sleepiness data is often imbalanced (more "alert" than "sleepy" states, or vice-versa), they explored several undersampling and oversampling techniques, including Exactly Balanced Bagging (EBB) and Roughly Balanced Bagging (RBB).
- Time Series Processing: To address the assumption of independent and identically distributed (i.i.d.) data in standard classifiers, they combined supervised classifiers with a Hidden Markov Model (HMM). The classifier predicts posterior probabilities, which are then used as likelihoods in the HMM's Viterbi decoding to produce a more natural and interpretable sequence of sleepiness predictions.
- Classification: Baseline classifiers such as Naive Bayes (NB), Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF), and a three-layer Neural Network (NN) were used.
Datasets
The paper introduces a newly collected dataset comprising synchronized respiration, accelerometer, and subjective sleepiness (KSS) data from 18 participants. This dataset was collected during typing and video-watching tasks, aiming to simulate natural daily living situations. The data was collected over approximately 3,420 minutes in total.
Key Insights and Ideas
- Respiration as a Rich Signal: The study highlights that respiratory patterns, often overlooked, contain valuable information about mental states like sleepiness, comparable to more complex physiological signals.
- Naturalistic Data is Crucial: The authors emphasize that sleepiness in daily life differs from driver drowsiness due to varying levels of resistance to sleep. Their experiment design, avoiding driving, aimed to capture this naturalistic sleepiness.
- HMM for Temporal Coherence: Integrating HMM with supervised classifiers significantly improves sleepiness prediction by enforcing temporal consistency, preventing unrealistic prediction sequences.
- Class Imbalance is a Real Hurdle: The study demonstrates that without proper handling, class imbalance in sleepiness data leads to biased predictions. Undersampling bagging methods like EBB and RBB proved effective.
- Interpersonal Variability: The F-measure varied significantly across subjects, suggesting that breathing patterns are highly individual. This points towards the need for personalized or adaptive sleepiness prediction models.
Key Results
- The best performing model, LR-EBB-HMM, achieved a mean F-measure of 0.7059. This is considered a significant improvement over baseline classifiers without class imbalance handling or HMM.
- Feature normalization was found to be crucial for improving the accuracy of posterior probability predictions, especially when combined with HMM.
- The combination of class imbalance handling techniques and HMM consistently improved the F-measure across all tested classifiers, with an average improvement of 0.1109.
- Visual inspection of prediction transitions (Figure 6) showed that LR-EBB-HMM could correctly predict sleepiness trends, unlike simpler models that were biased or produced unrealistic fluctuations.
Limitations and Future Work
- Limited Subject Count: The study involved only 18 subjects, which the authors acknowledge is a limitation. More data is needed for more accurate predictions and a better understanding of the physiological features and sleepiness association.
- Limited Task Variety: The data collection focused on typing and watching videos. Future work should investigate a broader range of daily activities to build a more robust model.
- Individual Adaptation: Due to significant interpersonal differences in breathing patterns, the authors suggest developing methodologies for individual adaptation of sleepiness prediction models.
- Weakly Supervised Learning: The high cost of collecting labeled data for sleepiness prediction suggests exploring weakly supervised learning and online learning approaches for future model development.
Summary updated