To quantify the degree of incoordination of gait in patients with multiple sclerosis (MS) using data from full-body motion capture device analyzed with Learning entropy (LE).

Variations of different walking parameters such as step length, cadence, step width, step angle, step time, joint angles, muscle force and many others can be estimated with inertial sensor-based systems or wearable systems. Perception Neuron (PN) is a commercial device containing 32 sensors each composed of a 3-axis gyroscope, accelerometer and magnetometer. Data obtained from the sensors have high statistical complexity. Assuming deterministic behavior of the system where newly obtained data follow governing laws  and therefore do not carry any novel information, we used Learning entropy (LE), which is a method of adaptive monitoring of dynamical systems, to quantify the degree of gait incoordination.

Seven ataxic patients with MS and 7 healthy controls were recorded using PN while walking. As the adaptive model input, the values recorded by 32 three-axis accelerometers were used. The predicted value was path length of the accelerometer to the next measurement. Normalized Least-mean-squares adaptive filter was used as the adaptive model. After estimating entropy from 32 sensors, its mean value and standard deviation were calculated. Support vector machine (SVM) was used for the classification of walking irregularities in patients with MS and healthy controls. The cross-validated SVM classifier was optimized using Bayesian optimization. The radial basis kernel function was selected for separation of the data. Parameters for balancing the error and margin width were optimized by quadratic programming.

Accuracy of discrimination between the two groups was 0.79.

Incoordination of gait in patients with multiple sclerosis can be quantified by Learning entropy. The prediction error of the model is probably due to the inability of patients to reproduce the optimal walking stereotype.