Abstract Body: Introduction: Peripheral Artery Disease (PAD) is a progressive vascular disorder impairing mobility, raising fall risk, and reducing quality of life. Early detection is key to preventing amputations and cardiovascular events. Current methods, often subjective, invasive, or costly, limit screening. Artificial intelligence and wearable technology enable non-invasive PAD detection via gait analysis. This study uses machine learning to classify PAD with gait stability and nonlinear dynamics.
Hypothesis: We hypothesize that machine learning models trained on Margin of Stability (MOS) and nonlinear gait features can classify PAD and Control groups, offering a non-invasive tool for early detection.
Goals: Objectives are to analyze gait stability differences between PAD and Control groups, quantify variability with nonlinear metrics, and develop a machine learning model for PAD classification.
Methods: Gait data from 139 PAD patients and 52 controls were collected using a motion capture system with passive markers to derive center of mass (COM). MOS was calculated in anterior-posterior (AP) and mediolateral (ML) directions. Spatiotemporal parameters (step duration, width, length, speed) and nonlinear dynamics (Lyapunov Exponent Wolf’s and Rosenstein’s [LyE_W, LyE_R], Sample Entropy [SpEn], Detrended Fluctuation Analysis [DFA], Correlation Dimension [CD], Approximate Entropy [ApEn]) were assessed. An Optimized Ensemble Model was trained on position-, velocity-, and acceleration-based features, evaluating MOS, spatiotemporal, and nonlinear contributions. Performance was measured by accuracy and feature importance.
Results: The Optimized Ensemble Model classified PAD versus Control groups. The position-based model achieved 92% validation and 92.6% test accuracy, velocity-based 85.3% and 85.2%, and acceleration-based 87.3% and 92.6%. Feature importance identified MOS_ML (0.0999), LyE_R (0.0953), LyE_W (0.0903) for position; Step time (0.7626), Step length (0.5028), CD (0.1920) for velocity; and Step time (0.1991), Step length (0.1938), ApEn (0.0730) for acceleration.
Conclusion: The Optimized Ensemble Model classifies PAD using gait stability and nonlinear dynamics. The position-based model (92.6% accuracy) excels, with MOS_ML and MOS_AP reflecting PAD stability deficits. LyE_R, LyE_W, and Step time show variability. Velocity (85.2%) and acceleration (92.6%) models highlight Step time and length, with ApEn capturing fluctuations. Future work will validate findings and test sensors.