Twenty-seven outpatients diagnosed with active, idiopathic unilateral BPPV of the posterior semicircular canal between the ages of 30 to 70 years (average 56.513.1), and twenty-seven healthy subjects between the ages of 25 to 70 years (average 56.110.8) were included in this study (Table 1). None of the healthy subjects had any medical history of neurological or orthopaedic conditions. According to the classification of patient's functional abilities by DHI scores, the BPPV patients were classified into three subgroups: mild stage (DHI = 0–30), moderate stage (DHI =31–60), and severe stage (DHI = 61–100).
The procedures of this study were approved and mandated by the institutional human research ethics committee of School of Biomedical Engineering, Shanghai Jiao Tong University (protocol number:201807), and conformed to the 1964 Helsinki Declaration. All subjects were fully informed of the study procedures, possible risks, privacy, and the freedom to withdraw. Informed consent was obtained from all participants.
2.2 Experiment setup
Level walking experiments were performed in the outpatient corridor of the neurology department at the Shanghai Ninth People’s Hospital. All subjects were instructed to walk at self-preferred speed along a 20 m walkway, during which their head was not allowed to turn, eyes looking straight ahead, and arms swinging naturally. A Timing Gait System (Brower, Draper, Utah, USA) was used to measure the walking duration of the middle 10 m of steady walk. The trial was defined as invalid if the standard deviation of the walking duration of each subject exceeded 5%. In this study, 6 valid trials were obtained for each subject. Two accelerometersDelsys, Inc., Boston, MA, USA)ere firmly strapped on the subject’s head and lower trunk at the third lumbar spinous process (L3) with belts (Fig. 1). Calibration was performed before each walking trail by placing it align with each orthogonal axis vertically to ensure the vertical acceleration is statically the ±1g value. Acceleration signals were captured by Delsys acquisition software (Delsys, Inc., Boston, MA, USA) and recorded at 148 Hz sampling rate in three orthogonal axes (VT, AP, and ML), respectively.
2.3 Data processing and gait variable calculation
The gravity component was first removed from the raw acceleration data and then filtered with a second-order Butterworth low pass filter with a cutoff frequency of 22 Hz. Five clinically relevant temporospatial variables and six variables reflecting walking stability were selected and calculated in Matlab (2019 a, the MathWorks, Inc., Natick, MA, USA).
Temporo-spatial variables: walking speed (m/s), walking distance (10 m) divided by the total time duration measured by timing gait system in the distance; step length(cm), walking distance (10 m) divided by the number of steps; cadence(steps/min), the number of vertical lower trunk acceleration peaks divided by the walking duration of each trial; step timing variability, SDs between successive gait cycles over an entire walking trial. Gait cycles were determined by the vertical lower trunk acceleration peaks.
Walking stability variables: Each variable in this part was calculated in the anteroposterior (AP), mediolateral (ML), and vertical (VT) axes. Acceleration root mean square (RMS), the dispersion of the measured acceleration signal relative to zero; Harmonic Ratio (HR), the ratio of even harmonics and odd harmonics of the measured acceleration signal, reflecting the gait smoothness and symmetry ; Step regularity (SR1), the amplitude of the first peak in the acceleration autocorrelation signal; Stride regularity (SR2), the amplitude of the second peak in the acceleration autocorrelation signal; Gait symmetry, the closeness of SR1/SR2 to 1.0 ; Gait variability, the width of the dominant peak in the power spectrum of the measured acceleration signal .
2.4 Statistical Analysis
All statistical analyses were performed using SPSS Release 22 (SPSS Inc., Chicago, IL, USA). All continuous variables were described with mean ± standard deviation. The normality test was performed using the Kolmogorov-Smirnov test and variables with positively skewed distributions were log10 transformed before inferential analysis. Walking stability variables were first adjusted to walking speed to remove the influence of gait speed [22,23]. One-way ANOVA was performed to test the differences of gait variables between BPPV patients and healthy controls.
2.5 Classification model of BPPV severity
A machine-learning based model was built for the classification of DHI subgroups of BPPV disease automatically.
Feature selection: To improve the performance of classification model, the gait variables that showed significant differences between BPPV patients and healthy controls were used as feature selection set for BPPV severity classification model. One-way ANOVA was used to further identify the model features by analyzing the significant differences of the gait variables among three disease stages of BPPV by DHI scores.
Model training: Support vector machine (SVM) with a linear kernel was used to build the model due to its good performance with high dimensional data, high signal to noise ratio , and it outperformed other machine learning algorithms, i.e. multi-layer perceptron and the k-nearest neighbors in training gait data .
Model validation: Repeated 5-fold cross-validation was performed to evaluate the model performance, meaning that the dataset was split into 5 subsets, where 4 subsets were used for training the model and the remaining subset was used as an independent validation set. This training and validation were repeated 5 times where each time a different independent validation set was used. The receiver operating characteristic (ROC) curve and accuracy (AUC) were used to evaluate the model performance in each fold.