Quantitative assessment of cerebellar ataxia, through automated limb functional tests

Abstract

Background: Cerebellar damage can often result in disabilities affecting the peripheral regions of the body. These include poor and inaccurate coordination, tremors and irregular movements that often manifest as disorders associated with balance, gait and speech. The severity assessment of Cerebellar ataxia (CA) is determined by expert opinion and is likely to be subjective in nature. This paper investigates automated versions of three commonly used tests: Finger to Nose test (FNT), test for upper limb Dysdiadochokinesia Test (DDK) and Heel to Shin Test (HST), in evaluating disability due to CA.

Methods: Limb movements associated with these tests are measured using Inertial Measurement Units (IMU) to capture the disability. Kinematic parameters such as acceleration, velocity and angle are considered in both time and frequency domain in three orthogonal axes to obtain relevant disability related information. The collective dominance in the data distributions of the underlying features were observed though the Principal Component Analysis (PCA). The dominant features were combined to substantiate the correlation with the expert clinical assessments through Linear Discriminant Analysis. Here, the Pearson correlation is used to examine the relationship between the objective assessments and the expert clinical scores while the performance was also verified by means of cross validation.

Results: The experimental results show that acceleration is a major feature in DDK and HST, whereas rotation is the main feature responsible for classification in FNT. Combining the features enhanced the correlations in each domain. The subject data was classified based on the severity information based on expert clinical scores.

Conclusion: For the predominantly translational movement in the upper limb FNT, the rotation captures disability and for the DDK test with predominantly rotational movements, the linear acceleration captures the disability but cannot be extended to the lower limb HST. The orthogonal direction manifestation of ataxia attributed to sensory measurements was determined for each test.

Trial registration: Human Research and Ethics Committee, Royal Victorian Eye and Ear Hospital, East Melbourne, Australia (HREC Reference Number: 11/994H/16).

Keywords: Diadochokinesia (DDK); Fast fourier transforms (FFT); Finger-to-nose (FNT); Heel shin test (HST); Principal component analysis (PCA).

Fig. 1

Schematic Representation of the Data Analysis: Feature Selection, Feature Extraction, Separation, Correlation and Classification

Fig. 2

Data analysis using BioKin^TM sensor: The data is transmitted wirelessly to the phone and then to the cloud storage. This is available for the data analysts

Fig. 3

Tests for evaluation of cerebellar ataxia disorder: Finger to Nose, test for upper limb Dysdiadochokinesia, and Heel to Shin. The 3 tests were performed by all the participants

Fig. 4

Resonant Frequency (RF) versus Magnitude (MR) from FFT analysis using features of high correlation. Figure 4a depicts Y-axis of gyroscope in FNT, Fig. 4b Z-axis of accelerometer in DDK, Fig. 4c depicts Z-axis of accelerometer in HST, and HST respectively

Fig. 5

Best separation using PCA analysis of kinematic parameters. Figures 5a, b, c depicts the best PCA separation on feature combination for the FNT, DDK, HST respectively

Fig. 6

Boxplot representing the feature separation with doctors score using Linear Discriminant Analysis Classifier. The 3 different classes of the classifier include controls, patients with low severity and patients with high severity of ataxia. The panel labels indicate the axes showing best performance of the classifier for the five kinematic parameters in this study. In FNT (Fig. 6b), angular acceleration features along Y-axis gave superior classification compared to the other parameters. In DDK, X and Z-axis features of acceleration (Fig. 6f) and Y-axis features of angle (Fig. 6j) shows highest discrimination. In HST, acceleration features of Y,Z axis (Fig. 6k) discriminated the cohort of patients and healthy subjects compared to the other parameters. These best outcomes from the LDA analysis is highlighted separately for acceleration and rotation

Fig. 7

Combination of FNT and DDK test (upper limb). The upper limb tests on combination gave good separation as given in figure (a) and classified based on severity values as given in Figure (b)

Fig. 8

ROC curve for the 3 tests. The AUC values calculated from ROC curve is found to be 0.7983 for (a) FNT, 0.9132 for (b) DDK and 0.8852 for (c) HST

Fig. 9

Cross Validation Parameters: Accuracy, Sensitivity and Specificity parameters for the 3 tests respectively