LiquidGAN for Handwriting-Based Detection and Severity Classification of Extrapyramidal Symptoms.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Extrapyramidal symptoms (EPS) are motor side effects of antipsychotic medications.
• EPS manifest as measurable changes in handwriting, affecting spatial and temporal writing characteristics.
• Existing models struggle to capture fine-grained variations in EPS severity.

Purpose of the Study:

• To develop a novel model, Liquid Generative Adversarial Network (LiquidGAN), for analyzing handwriting changes associated with EPS.
• To compare the performance of LiquidGAN against conventional models like ResNet50 for EPS classification.
• To assess LiquidGAN's capability in generating realistic synthetic handwriting data for improved analysis.

Main Methods:

• Collected handwriting data (Archimedean spirals) from patients with EPS and healthy controls.
• Utilized a Liquid Generative Adversarial Network (LiquidGAN) combining liquid neural networks and GANs.
• Employed ResNet50 for baseline binary and multiclass classification of handwriting patterns.

Main Results:

• LiquidGAN achieved 97% accuracy and 98% precision in binary EPS classification.
• LiquidGAN significantly outperformed ResNet50 in multiclass classification (70% accuracy) for EPS severity.
• LiquidGAN generated high-quality synthetic handwriting data, improving classification and outperforming StyleGAN.

Conclusions:

• Handwriting analysis using LiquidGAN provides an effective, non-invasive method for monitoring EPS.
• LiquidGAN captures subtle handwriting variations linked to EPS severity, outperforming conventional deep learning models.
• Dynamic generative learning with handwriting biomarkers shows promise for clinical EPS monitoring.