Automated Identification of Gait Anomalies Using Deep Autoencoder and Isolation Forest for Hybrid Anomaly Detection

Human gait analysis plays a vital role in assessing locomotor function, postural stability, and early detection of motor impairments. This study proposes an unsupervised hybrid anomaly detection framework that integrates PCA and Isolation Forest (IF) to automatically identify abnormal gait patterns using a Multivariate Biomechanical Dataset (MGAD) containing 5,000 gait samples. PCA was utilized to reduce dimensionality and compress correlated gait features while retaining 95.1% of the total variance, thereby preserving essential biomechanical information. The reconstruction errors obtained from PCA were subsequently analyzed using Isolation Forest to isolate anomalous gait instances. Experimental results demonstrate that the hybrid PCA–IF model effectively differentiates between normal and abnormal gait behaviors, achieving an ROC-AUC of 0.912 and an F1-score of 0.866, indicating strong discriminative capability and model stability. Feature-level reconstruction analysis revealed that stance phase duration, step length, and stride length are the most influential determinants of gait irregularities, aligning with established clinical findings in gait biomechanics. The proposed framework is computationally efficient, interpretable, and fully unsupervised, making it suitable for real-time clinical assessment, rehabilitation monitoring, and wearable healthcare applications. These findings highlight the potential of hybrid statistical–machine learning models in advancing automated gait diagnostics and intelligent mobility analytics.