TL;DR
This paper investigates the effectiveness of single wearable sensors (accelerometer, gyroscope, sEMG) for recognizing daily activities in both able-bodied individuals and stroke survivors. They found that an accelerometer, particularly when placed on the extensor carpi ulnaris, offers the best performance, even outperforming sensor combinations in some scenarios, and explored the feasibility of using a model trained on healthy individuals to recognize activities of stroke survivors.
What Problem Does This Paper Address?
Human Activity Recognition (HAR) is crucial for various applications, including health monitoring and rehabilitation. While multi-sensor systems offer high accuracy, they are often costly, inconvenient, and computationally intensive. For stroke survivors, who often experience motor impairments, accurate and unobtrusive HAR is particularly important for assessing recovery and supporting daily living. However, existing research has not comprehensively compared the performance of different single sensors or explored the optimal placement for these sensors, especially in the context of stroke survivors. This paper aims to fill this gap by systematically evaluating single sensors and their placement for HAR, considering both healthy individuals and stroke survivors.
Methods
The study employed a comparative approach to evaluate the performance of three types of wearable sensors: accelerometers (ACC), gyroscopes (GYRO), and surface electromyography (sEMG). These sensors were integrated into Delsys wireless modules, with 12 modules placed at various key positions on the upper and lower limbs.
The core methodology involved:
- Data Collection: Four daily activities (walking, tooth brushing, face washing, drinking) were recorded from 23 participants (14 healthy, 9 stroke survivors) using the 12 sensor modules.
- Data Pre-processing: Raw sensor data was filtered (low-pass for ACC/GYRO, band-pass for sEMG) and normalized using z-score normalization.
- Feature Extraction: Time-domain features were extracted from the processed signals. For ACC, 45 features were extracted; for GYRO, 39; and for sEMG, 13. Combinations of sensors also yielded a larger feature set.
- Imbalanced Data Handling: Due to stroke survivors sometimes ending trials early, leading to imbalanced datasets, the Synthetic Minority Over-sampling Technique (SMOTE) was used to balance the training data.
- Classification: The Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel was chosen as the primary classifier due to its superior performance compared to other tested classifiers (KNN, DT, AdaBoost, RF, LDA, ANN, GNB, Linear SVM).
- Validation: Leave-One-Subject-Out Cross-Validation (LOSOCV) was employed to ensure objective and realistic performance evaluation.
- Two Scenarios:
- Scenario 1: Evaluating HAR performance within three groups: stroke survivors, healthy subjects, and all subjects combined. This scenario aimed to identify the best sensor and placement for each group.
- Scenario 2: Training a HAR model on healthy subjects and testing it on stroke survivors to assess the feasibility of a pre-trained model for practical applications.
- Feature Selection: The Minimum-Redundancy Maximum-Relevancy (mRMR) algorithm was used to select the most informative features.
Datasets
The study utilized a dataset collected from 23 participants (14 healthy individuals and 9 stroke survivors) performing four specific daily activities. This dataset was collected specifically for this research and is not publicly available due to privacy concerns. The study did not introduce a new benchmark dataset but rather used its collected data to evaluate existing HAR techniques and sensor modalities.
Key Insights and Ideas
- Accelerometer Dominance: The accelerometer consistently outperformed gyroscopes and sEMG sensors for recognizing daily activities, even when used as a single sensor. This suggests that simple motion detection is highly effective for these activities.
- Optimal Sensor Placement Matters: The location of the sensor significantly impacts performance. The extensor carpi ulnaris (US6) emerged as a consistently high-performing location for accelerometers across different groups.
- Single Sensor Sufficiency: For many daily activities, a single accelerometer can achieve performance comparable to, or even exceeding, sensor combinations, significantly reducing system complexity and cost.
- Pre-trained Models for Stroke Survivors: A model trained on healthy individuals can provide a reasonable starting point for recognizing stroke survivor activities, though performance is lower than when stroke survivor data is included in training. This highlights the potential for transfer learning but also the need for some adaptation.
- sEMG Challenges: sEMG proved to be the least effective sensor, likely due to the complexity of muscle signals, individual variations, and the impact of motor impairments in stroke survivors, which can lead to abnormal and inconsistent muscle activation patterns.
Key Results
- Highest Accuracy (Scenario 1): The accelerometer placed on the extensor carpi ulnaris (US6) achieved the highest accuracy of 95.84 ± 1.75% for the group of all subjects. For healthy subjects, the accelerometer on the brachioradialis (US5) yielded the highest accuracy (96.43 ± 2.35%), while US6 was best for stroke survivors (94.05 ± 3.1%).
- Scenario 2 Performance: When using a model trained on healthy subjects to classify stroke survivor activities, the highest accuracy achieved was 77.89 ± 4.81% using the accelerometer on the extensor carpi ulnaris (US6).
- Sensor Comparison: Across all groups and scenarios, the accelerometer generally provided the best performance among single sensors. Sensor combinations offered marginal improvements in some cases but were not always significantly better than the best single accelerometer placement.
- Statistical Significance: Statistical analysis confirmed significant differences in performance between sensor types and combinations within groups, with post-hoc analysis revealing specific relationships.
Limitations and Future Work
- Limited Activities: The study focused on only four specific daily activities. Future work should include a wider range of activities, such as lying, standing, and stair climbing, to assess generalization.
- Stroke Survivor Heterogeneity: The recruited stroke survivors were primarily at Brunnstrom stages IV and V. Future research should include a more diverse range of stroke severities to better assess generalization across different levels of impairment.
- Data Scarcity: Obtaining sufficient data from stroke survivors is challenging due to their physical and mental burden. Future work could explore more advanced techniques to leverage limited stroke survivor data, potentially in combination with larger healthy subject datasets.
- Real-World Application: While the study explored a pre-trained model for stroke survivors, further validation in real-world, home-based scenarios is needed.
- Stroke Impairment Assessment: The authors suggest that the involved activities could be used to evaluate stroke impairment without needing to record activity sequences, which warrants further investigation.
Summary updated