Smart Wearable Detects Falls in Elderly With 96.7% Accuracy Using AI and LoRa

Falls are the second leading cause of unintentional trauma death globally, killing an estimated 684,000 people annually, with adults over 60 at greatest risk. Nearly half of older adults who fall cannot rise unassisted, making rapid detection and alerting essential. Yet existing wearable fall detectors often sacrifice battery life for accuracy, or rely on short-range Bluetooth that requires a nearby smartphone—limiting real-world reliability.

This study from researchers at Universidad de Málaga and Universidad de Investigación y Desarrollo (Colombia) proposes a hybrid wearable system that addresses all three core limitations simultaneously: detection accuracy, energy efficiency, and communication range. The device integrates an accelerometer sampling at 20 Hz with a two-stage detection pipeline. In stage one, a lightweight threshold-based model monitors continuous acceleration magnitude and triggers the system only when a predefined limit is exceeded—keeping the processor and LoRa radio in deep sleep otherwise. In stage two, a 4-second windowed acceleration sample is transmitted via LoRa to a remote server, where a CNN-LSTM deep learning model performs final fall/non-fall classification. The CNN extracts spatial features from the accelerometer signal while the LSTM captures temporal movement patterns, together reducing false positives that simpler threshold models generate.

The prototype achieved 96.67% sensitivity (detection rate for actual falls) and 100% specificity (no misclassification of Activities of Daily Living as falls) during experimental testing with simulated fall scenarios. The ultra-low-power design extended battery autonomy to 178 hours of continuous monitoring—significantly longer than many competing systems. LoRa's LPWAN architecture enables kilometer-scale transmission without cellular infrastructure, operating on unlicensed frequency bands to eliminate recurring network costs and smartphone dependency.

The system outperforms several prior LoRa-based and NB-IoT-based fall detectors reviewed in the paper. For context, comparable systems achieved accuracies ranging from 89.2% to 96.93%, often with shorter battery life or limited communication range. The hybrid algorithm design—offloading computationally expensive deep learning to a remote server rather than the wearable itself—is a key architectural insight enabling both high accuracy and low power draw simultaneously.

Important caveats apply: fall scenarios were simulated rather than observed in real older adults in naturalistic settings, which may overestimate real-world performance. The dataset size and subject diversity are not fully detailed in the available text, and the system has not yet been validated in clinical or long-term home monitoring trials. Additionally, LoRa gateway infrastructure must be available in deployment environments, which may limit applicability in some rural or indoor settings without existing LPWAN coverage.