Differentiation benign paroxysmal positional vertigo from vestibular migraine via walking stability analysis: a cross- sectional and machine-learning study

Abstract Background Current clinical methods of distinguishing benign paroxysmal positional vertigo (BPPV) and vestibular migraine (VM) are primarily based on medical history and characteristics of nystagmus lacking objective and precise techniques. Methods In this study, we developed a machine-learning-based approach to distinguish BPPV and VM via gait performance and walking stability analysis. 16 BPPV patients, 16 VM patients and 16 healthy participants performed 10m level over-ground walking trials at self-preferred speed while wearing accelerometers on the head, the trunk and the ankles. Gait spatial-temporal and walking stability characteristics, including acceleration root mean square (RMS), harmonic ratio (HR), amplitude variability (AV), step/stride regularity and gait symmetry, were statistically analyzed within three groups. Ten kinds of individual learning and ensemble learning models were trained to classify participants into the BPPV group, VM group and HC group, based on walking stability characteristics. Results Results showed that the walking speeds of VM and BPPV patients were lower than those of healthy participants. Head acceleration RMS in the AP axis, step regularity in the mediolateral (ML) and head vertical (VT) axes decreased in VM and BPPV patients compared with healthy participants. Simultaneously, acceleration RMS in the ML axis of the head, AV in the VT axes at the head and trunk and HR in the VT axis of both head and trunk showed differential significance between VM and BPPV patients. The random forests (RF) model showed better classification performance with 83.9% accuracy and 0.854 AUC. Conclusions This study demonstrates the feasibility of distinguishing VM and BPPV based on walking stability parameters and machine learning models.