LSTM-CNN: An efficient diagnostic network for Parkinson's disease utilizing dynamic handwriting analysis

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Dynamic handwriting analysis offers a non-invasive method for early Parkinson's disease (PD) diagnosis.
• Quantifying PD dysgraphia requires analyzing subtle signal variations, with potential importance in local signal segments.
• Previous methods may overlook the significance of local signal segments in handwriting analysis.

Purpose of the Study:

• To propose a lightweight network architecture for analyzing dynamic handwriting signal segments.
• To develop an efficient method for quantifying PD dysgraphia and aiding early PD diagnosis.
• To present visual diagnostic results for improved clinical utility.

Main Methods:

• Investigated time-dependent patterns in local handwriting signal representations.
• Segmented handwriting signals into fixed-length sequential segments.
• Designed a compact 1D hybrid network for feature extraction and classification, using majority voting for final diagnosis.

Main Results:

• Achieved high diagnostic performance on DraWritePD (96.2% accuracy) and PaHaW (90.7% accuracy) datasets.
• The network is lightweight with 0.084M parameters and 0.59M FLOPs.
• Demonstrated near real-time CPU inference performance (0.106-0.220s per signal).

Conclusions:

• The proposed method effectively and efficiently quantifies dysgraphia for precise PD diagnosis.
• The study systematically demonstrates the method's effectiveness through extensive experiments.
• The lightweight design and high performance offer a practical diagnostic tool.