GaitGuardian: Real-Time IMU Processing and Visual Monitoring Using Machine Learning (ML) for Advanced Parkinson's Disease Patients (II)

Overview

Over 10 million people live with Parkinson's Disease (PD), an incurable movement disorder that disrupts daily life. Freezing of Gait (FoG) affects 80% of advanced patients, causing sudden immobility and falls. Accordingly, PD patients often look down while walking, needing navigation assistance. Additionally, over 75% of patients undergo rest tremors (uncontrolled shaking), disrupting motor tasks. Hence, this work proposes GaitGuardian — a holistic machine learning (ML) solution for advanced PD patients, including FoG prediction, fall and tremor detection, and visual monitoring. First, the FoG prediction algorithm uses accelerometer data to flag a “pre-FoG” state (2s before FoG). A hybrid CNN + biLSTM model enhanced with dual-attention captures spatial and temporal features, achieving high sensitivity and specificity. Fall detection runs an ML ensemble on IMU data, nearing perfect accuracy. Tremor detection uses a lightweight 1D CNN trained on wrist-worn IMU data. Finally, visual monitoring pairs an object + depth detection backbone using transformer and convolutional modules (cloud-based) with a multimodal large language model for text output. ML techniques like oversampling, feature extraction, low-pass filtering, and tuning are applied to optimize accuracy and computation. For physical construction, a 6-DoF IMU connected to an ESP32-S3 is cased in PLA, using USB-C for power and data. The device mounts on the trunk while walking, or on the wrist for tremor remediation via haptics. Finally, a 2 MP Arducam is placed on the user’s forehead, using BLE-5 for data transmission (via ESP32). Overall, GaitGuardian can predict FoG, detect falls, remediate tremors, and monitor surroundings while outperforming existing models, demonstrating its beneficial impact on PD patients.