A COMPARATIVE STUDY OF DEEP LEARNING ARCHITECTURES FOR REAL-TIME FALL DETECTION USING WEARABLE SENSOR DATA

Abstract

Falls remain a leading cause of injury among elderly and mobility-impaired individuals. This study presents a comparative assessment of deep learning architectures for real-time fall detection using wearable sensor data. The proposed system integrates data preprocessing, model training, and real-time inference using CNN, LSTM, BiLSTM, and Transformer models. The research focuses on optimizing performance and latency to ensure reliable operation on embedded and wearable platforms. The CNN-LSTM amalgamation yielded enhanced performance because these models excel in extracting both spatial information from multivariate time-series data together with temporal feature modeling. Evaluations made within the scope of the CNN-LSTM method showed that a rate of 97.3% could be achieved. This result stood out as one of the highest performance levels reported in literature.  Findings contribute toward building robust, real-world fall detection systems with practical deployment potential in healthcare monitoring and assisted living environments.