To enable prompt intervention and lower the risk of serious injuries, especially among the elderly, accurate and timely fall detection is essential. We suggest a deep learning-based framework that makes use of data from a custom made, wrist-worn IoT device built on the ESP32 platform in order to address this pressing need. The gadget continuously records multi-sensor data from four people carrying out a predetermined set of 12 tasks, including readings from an accelerometer, gyroscope, and heart rate monitor (BPM). Six fall scenarios and six daily living tasks are included in these activities, which adhere to standard protocols. Rigid feature extraction across statistical, temporal, and frequency domains produced a manually annotated dataset. An artificial neural network (ANN) that can accurately differentiate fall events from non-fall scenarios under a variety of movement conditions was trained using the extracted features. Despite utilizing a comparatively small dataset collected from four young subjects, the experimental results show that the system achieves a high accuracy rate of 97%. Although this preliminary validation is encouraging, more research involving older populations and more varied datasets is required to determine clinical generalizability. Significantly lowering false alarms and highlighting its potential for use in realistic healthcare and emergency response settings, the combination of acceleration, angular velocity, and BPM data produced better performance than individual sensor models.