Unsupervised Learning as a Complement to Convolutional Neural Network Classification in the Analysis of Saccadic Eye Movement in Spino-Cerebellar Ataxia Type 2

Abstract

This paper aims at assessing spino-cerebellar type 2 ataxia by classifying electrooculography records into registers corresponding to healthy, presymptomatic and ill individuals. The primary used technique is the convolutional neural network applied to the time series of eye movements, called saccades. The problem is exceptionally hard, though, because the recorded saccadic movements for presymptomatic cases often do not substantially differ from those of healthy individuals. Precisely this distinction is of the utmost clinical importance, since early intervention on presymptomatic patients can ameliorate symptoms or at least slow their progression. Yet, each register contains a number of saccades that, although not consistent with the current label, have not been considered indicative of another class by the examining physicians. As a consequence, an unsupervised learning mechanism may be more suitable to handle this form of misclassification. Thus, our proposal introduces the k-means approach and the SOM method, as complementary techniques to analyse the time series. The three techniques operating in tandem lead to a well performing solution to this diagnosis problem.

A hybrid unsupervised—Deep learning tandem for electrooculography time series analysis

“…A preliminary study in [4] outlined the separate findings of a trio of techniques for both supervised and unsupervised analysis of the previous version of the saccadic data. A temporal CNN from the classification side, and K-means and SOM from the clustering perspective, each discovered different patterns in the shapes of the saccades.…”

“…The corresponding results for validation are of 73.23% and 60.82%, respectively. Alternatively, a single CNN architecture was tried for the same problem, as before in [4], but its results were significantly worse than those of the CNN-LSTM approach.…”

“…Both aspects pose a serious problem for machine learning methodologies that strive to discover the relationship between shape and diagnosis class. As such, their accuracy of discrimination cannot go beyond 66% [3,4], no matter the techniques employed.…”

Automated Detection of Presymptomatic Conditions in Spinocerebellar Ataxia Type 2 Using Monte Carlo Dropout and Deep Neural Network Techniques with Electrooculogram Signals

“…Segments from the velocity profiles were grouped into two categories using the k-means algorithm, and segments corresponding to the group with high velocity were considered as saccades. Each saccade has a length of 192 samples at regular time steps with normalized amplitude in the interval [−0.5, 0.5] (see [2] for a detailed description of the acquisition and preprocessing stages). In total, there are 5953 saccades collected for all the individuals that had undergone examination.…”

“…However, the performance has improved only when a supporting unsupervised technique (viz. self-organizing maps (SOM)) was used in parallel with DL [2] or to clean the data before the application of a deep neural network (DNN) [3]. In this context, uncertainty quantification (UQ) may be a promising mechanism to include in the DNN, thus allowing to better differentiate the presymptomatic condition by addressing both data ambiguity and model behavior.…”