Real-time fall prediction and protection systems can significantly reduce fall-related injury risks while enhancing independence, physical well-being, and mental health of elderly individuals living alone. To improve fall prediction algorithm performance, specifically recognition accuracy, recall rate, and specificity, while minimizing both fall misclassification errors and airbag deployment time, this study proposes a multi-threshold fall prediction algorithm based on support vector machines (SVM), integrated with an airbag protection system. Motion data are first collected through a waist-worn acceleration sensor. Then, the SVM algorithm determines optimal thresholds for acceleration, velocity, and posture angle to differentiate falls from activities of daily living (ADLs). Finally, the optimized algorithm is deployed on a microcontroller to enable real-time fall prediction and trigger the airbag system. Experimental results show that the system achieves 97.3% accuracy, 99% recall and 96.1% specificity in fall recognition, with an average airbag inflation time of 350.4 ms. These metrics confirm both high prediction reliability and rapid protective response, validating the system's effectiveness for real-time fall prediction and protection.