Continuous Freezing of Gait Detection With TinyML on Embedded Systems

This paper presents the deployment of quantized Freezing of Gait (FOG) detection models on resource-constrained devices like the Arduino Nano BLE 33. Conventional machine learning (ML) models tend to be too large and computationally intensive for real-time, on-device use, often overlooking critical factors like inference time and power consumption. In contrast, our approach employs a dense neural network (DNN) quantized from float32 to int8 post training, reducing the model size to 15-20 kB while maintaining high accuracy and computational efficiency. The integration of multi-modal sensor data, such as, acceleration (ACC), electromyography (EMG), and skin conductance (SC) is important, as it helps identify which features most effectively contribute to accurate FOG detection while optimizing power and inference time. This work demonstrates that the model combining 36 features slightly increases inference time by 0.6 ms compared to the same model with only 6 features, but improves detection accuracy (89%), sensitivity (88%), specificity (89%), F1 score (88%), and AUC (0.94). The model’s power consumption stands at 5 mA during inference, making it highly efficient for long-term use. Minimizing inference time is critical for wearables, as it directly impacts power consumption during each cycle, allowing for more efficient duty cycling and extended battery life. By quantizing the model and examining trade-offs between feature richness, inference time, and power usage, we demonstrate the effectiveness of this approach for real-time, low-power FOG detection on wearable devices, making it a practical solution for continuous health monitoring.