A novel ensemble model for fall detection: leveraging CNN and BiLSTM with channel and temporal attention
Falls are a serious public health concern in a society where the elderly population is growing and requires prompt medical attention. Despite the proliferation of machine learning and deep learning algorithms for fall detection, their efficacy remains hampered by resilience, robustness, and adaptabi...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Taylor & Francis Group
2025-04-01
|
| Series: | Automatika |
| Subjects: | |
| Online Access: | https://www.tandfonline.com/doi/10.1080/00051144.2025.2450553 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Falls are a serious public health concern in a society where the elderly population is growing and requires prompt medical attention. Despite the proliferation of machine learning and deep learning algorithms for fall detection, their efficacy remains hampered by resilience, robustness, and adaptability challenges across varied input scenarios. When using models that utilize multiple sensors, giving equal importance to each sensor can lead to errors because some activities may appear similar. To address this issue, researchers propose integrating attention mechanisms, which help prioritize important information from the sensors and reduce the impact of over lapping activity patterns. These challenges limit their practical implementation in wearable systems. To address these limitations, this study introduces an innovative attention-based ensemble model for fall detection; by integrating a convolutional neural network with channel attention and a bidirectional long short-term memory with temporal attention, the model prioritizes relevant information within time series data. The channel attention module uncovers interrelationships between variables. Meanwhile, the temporal attention module captures associations within the sensor data’s temporal dimension, allowing the model to focus on critical features and enhance performance. The experimental findings reveal impressive classification accuracies of 97.93% and 98.99% on the KFall and SisFall datasets, respectively. |
|---|---|
| ISSN: | 0005-1144 1848-3380 |