Edge-based freezing of gait recognition in Parkinson's disease

Freezing of gait (FoG) stands as one of the most debilitating symptoms of Parkinson's disease (PD), occurring in more than half of patients with advanced PD. This condition manifests as a sudden blockage, significantly reducing the patients’ quality of life. To improve gait and ameliorate FoG, cueing strategies involving audio, visual, or tactile stimulation have been evaluated. In particular, on-demand systems that can automatically detect FoG and administer cueing have emerged as promising solutions. In response, several wearable sensors and machine learning-based approaches have been proposed for accurate FoG recognition. However, existing techniques suffer from several critical challenges, notably suboptimal performance, and limitations for real-time operation and edge deployment. Addressing these issues, this study presents a groundbreaking advancement in real-time edge-based FoG recognition utilizing convolutional neural networks (CNN). We designed an optimized model, rigorously evaluating it across 62 PD patients using a cutting-edge reference dataset, achieving an F1-score of 92% and an area under the curve of 0.97. Further testing on an external dataset resulted in consistent detection performance, while a lower specificity was observed. The CNN implementation on a cost-effective processing device resulted in a 1 ms inference time and required only 6.3 KB of random access memory (RAM) and 37.8 Kb of Flash memory, meeting real-time demands and enhancing clinical applicability.