Related Experiment Video
Updated: Dec 14, 2025
06:12
Recording Horizontal Saccade Performances Accurately in Neurological Patients Using Electro-oculogram
Published on: March 13, 2018
11.0K
A hybrid unsupervised-Deep learning tandem for electrooculography time series analysis
Ruxandra Stoean1, Catalin Stoean1, Roberto Becerra-García2
1University of Craiova, Craiova, Romania.
Plos One
|July 22, 2020
Summary
This study improves early detection of spinocerebellar ataxia type 2 by combining clustering with deep learning. This method successfully identified half of presymptomatic cases, crucial for timely intervention.
Area of Science:
- Computational neuroscience
- Medical data analysis
- Machine learning in healthcare
Background:
- Deep learning methods struggle with complex medical data, particularly distinguishing between similar patient classifications.
- Spinocerebellar ataxia type 2 (SCA2) diagnosis presents challenges due to overlapping sample characteristics in electrooculography (EOG) tests.
Purpose of the Study:
- To develop an improved computational method for diagnosing spinocerebellar ataxia type 2 using electrooculography data.
- To enhance the identification of presymptomatic cases, which is critical for early intervention.
Main Methods:
- A hybrid approach combining self-organizing maps (SOM) for pre-processing clustering and a deep learning model (convolutional and long short-term memory networks) for classification.
- Data relabeling based on shape similarity identified through SOM clustering before deep learning training.
Main Results:
- The dual methodology achieved 78.24% accuracy in discriminating three classes of EOG test results for SCA2.
- The combined approach significantly improved the detection of presymptomatic SCA2 cases, identifying 50% correctly, unlike single deep learning models.
Conclusions:
- The integration of clustering with deep learning offers a more effective strategy for medical data classification, particularly for challenging datasets like EOG in SCA2.
- Prioritizing the identification of presymptomatic cases over marginal overall accuracy gains is vital for clinical utility in neurodegenerative disease diagnosis.
