MIEF-Net: multimodal image-enhanced fusion network for intelligent fall risk prediction

Falls among older adults are a major health concern, often resulting in severe injuries and diminished quality of life. Early prediction of fall risk and timely intervention can significantly reduce occurrence of negative health outcomes. To date, data-driven approaches that integrate Inertial Measurement Unit (IMU) devices have received growing attention in automated fall risk assessment due to their portability, ease of use, and reduced dependence on healthcare professionals. This study proposes a novel image-enhanced dual-stream deep learning approach to fall risk prediction through IMUs based gait analysis. We present an innovative fusion of spatial-temporal representations by transforming raw IMU signals into Gramian Angular Field (GAF), spectrogram, and Markov Transition Field (MTF) images. These complementary modalities respectively capture the inherent temporal correlations, frequency patterns, and state transition dynamics in gait. The proposed approach integrates recurrent neural networks (RNN) for sequential modeling with convolutional neural networks (CNN) for spatial feature extraction, enhanced by a Transformer-based multi-head attention mechanism for adaptive modality fusion. Evaluated on a clinical data collected from field study, the proposed approach achieves superior performance (Accuracy: 0.9712, Sensitivity: 0.9697, F1-score: 0.9708) and significantly outperforms the state-of-the-art models. The dual-stream design enables detection of subtle gait abnormalities that conventional methods overlook. This work advances preventive geriatric care by demonstrating how wearable sensor-based multi-modal fusion enhances fall risk prediction accuracy while maintaining clinical practicality.