تشخيص زمين خوردن سالمندان تحت مراقبت از راه دور در افزاره‌هاي پوشيدني خود توان

Falls represent one of the most critical health hazards for the elderly population, often leading to prolonged hospitalization, permanent disability, an‎d increased mortality due to delays in receiving medical assistance. While wearable fall detection systems enable continuous health monitoring, their practical efficacy is significantly hindered by limited battery capacity an‎d the need for frequent recharging, which is particularly problematic for elderly users. To overcome these limitations, this thesis proposes an energy-sustainable fall detection framework based on deep learning. The framework was developed using the public SisFall dataset containing data from 38 participants performing 19 activities of daily living an‎d 15 simulated fall scenarios. A comprehensive preprocessing pipeline was implemented including filtering, sliding window segmentation, an‎d normalization. Thirty-seven kinematic features were extracted including time-domain statistics, Signal Magnitude Area (SMA), an‎d Jerk. In the initial stage, a Ran‎dom Forest classifier was implemented as a baseline model, achieving 89.69% accuracy but only 76.45% sensitivity. Subsequently, the framework was enhanced by developing a Bidirectional Long Short-Term Memory (Bi-LSTM) neural network. To address the severe class imbalance characteristic of fall detection datasets, an optimized version incorporating three Bi-LSTM layers with 256 units, an 8-Head Self-Attention mechanism, an‎d the Focal Loss function was proposed. The final optimized model achieved outstan‎ding performance with accuracy of 97.74%, sensitivity of 95.44%, specificity of 98.83%, F1-Score of 96.45%, an‎d AUC-ROC of 99.65%. This represents an 18.99% improvement in sensitivity an‎d a 13.06% improvement in F1-Score compared to the baseline model. The confusion matrix analysis revealed a remarkable reduction in false negatives from 380 to 72, demonstrating the modelʹs ability to reliably detect falls. The results demonstrate that the proposed framework provides a robust, autonomous, an‎d deployable solution for long-term elderly health monitoring an‎d telemedicine applications, with significant improvements in sensitivity an‎d overall performance.

مينا اميري

احمدرضا تابش

محمد داورپناه جزي , الهام محمودزاده