Data Processing using Machine Learning techniques for Fall Detection and Prevention

Falls constitute a major global health and occupational safety crisis, resulting in millions of injuries and fatalities annually. Traditional fall safety measures are reactive, relying on kinematic data to detect an event only after it has occurred, which fails to enable timely injury mitigation. This PhD thesis develops a comprehensive, machine learning framework that transitions fall safety from reactive reporting to continuous, proactive prediction, leveraging multimodal biosensing and personalized data processing techniques. The research establishes a three-part framework: Reactive Detection, Proactive Impact Prediction, and Predictive Risk Estimation. First, a Deep Neural Network (DNN) achieved a high True Negative Rate of 94.4% and a competitive True Positive Rate of 82.6% on the SisFall dataset, confirming robust fall detection capabilities while highlighting the limitations of kinematic data alone, particularly in detecting subtle falls. Second, to enable injury mitigation systems, such as wearable airbags, the system integrated Proactive Impact Prediction. A novel Kolmogorov–Arnold Network (Fall-KAN) was introduced, achieving a superior mean Root Mean Squared Error of approximately 159 milliseconds for Time-to-Impact estimation, demonstrating the feasibility of millisecondscale prediction. Alternatively, a Deep Neural Network architecture trained on an artificially enriched dataset generated using the Inverted Pendulum was tested, which forecasted the fall trajectory and provided a crucial preventive time window of approximately 200 milliseconds. Third, the thesis addresses the root causes of falls through Predictive Risk Estimation, utilizing multimodal physiological data collected during Activities of Daily Living and highstress Virtual Reality fall simulations. Crucially, analysis revealed that generalized Machine Learning models failed to reliably classify fall risk across different subjects due to high inter-subject variability. Consequently, the framework shifted to a personalized classification approach, which achieved highly robust performance, attaining a balanced accuracy of up to 90.39% on the test set for binary (Low vs. High) risk classification. In conclusion, this work validates a holistic framework capable of reliable fall detection, high-speed impact forecasting, and robust personalized risk assessment. By fusing advanced deep learning architectures with individual physiological signatures, this research provides the essential scientific foundation for the next generation of truly proactive, life-saving fall prevention and injury mitigation technologies.