Differentiation of Neurodegenerative Diseases by Dynamic Analysis of Gait Pattern and ‎Feature-level Fusion Approaches

In recent years, some studies have examined the gait patterns of neurodegenerative diseases utilizing signal processing techniques and machine learning algorithms. The aim of this study was to provide an automated system for distinguishing Huntington's disease, amyotrophic lateral sclerosis (ALS), and Parkinson's disease from healthy control group using dynamic analysis of gait pattern (more precisely, stride time). In addition, we examined the effect of fusion of features obtained from the left and right feet.

First, polar-based measures were extracted from lagged Poincaré maps. The optimal latency of the map was estimated using the mutual information algorithm. Then, five feature-level fusion strategies were presented. The classification was performed using the feed-forward neural network; while the effect of changing the network parameter was also investigated. The proposed system was evaluated using the data available in the Physionet database, which includes 16 records of the control group (14 females and 2 males; 20-74 years), 20 records of Huntington's disease (14 females and 6 males; 29-71 years), 13 records of ALS (3 women and 10 men; 36-70 years) and 15 records of Parkinson's disease (5 women and 10 men; 44-80 years).

Using the fourth fusion strategy, the maximum accuracy of 93.47% was obtained in separating the control and Huntington groups. Applying the second fusion algorithm, the control/Huntington and control/Parkinson groups were separated with the accuracy rate of 92.92% and 91.93%, respectively. The highest accuracy of the first fusion algorithm was 91.72% in classifying the control group and ALS. The third fusion algorithm was able to provide a 91.13% classification accuracy in separating the control and Huntington groups. The performance of the algorithm in separating patient groups was weaker.

The proposed system performed well compared to previously published algorithms. Further studies on intelligent classification algorithms and the development of the suggested method could pave the way for preclinical diagnosis of neurodegenerative diseases.