Chen-Chang Shih, Syaun-Huei Lin, Wei-Jang Yen, Guang-Chih Cheng, Whai-En Chen, Chih-Chung Shiao
Falls in the elderly represent a critical public health crisis, necessitating a shift from reactive to proactive management. This review evaluates artificial intelligence (AI) across four primary modalities: wearable sensors, vision-based systems, ambient devices, and natural language processing and large language models applied to electronic health records. Machine learning and deep learning algorithms enable real-time detection, near-fall identification, and predictive risk stratification. Hybrid edge-cloud frameworks and multimodal fusion enhance accuracy and scalability. However, challenges include the "simulation gap" in training data, "black box" interpretability, and privacy concerns, "alarm fatigue" among nursing staff, and "class imbalance." Future directions emphasize hybrid systems, integrating multiple sensor streams with edge-cloud computing, to enhance ecological validity and responsiveness. Ultimately, seamless integration of interpretable AI into clinical workflows offers a transformative path toward personalized prevention, reducing fall-related morbidity and empowering elderly independence.