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

Dynamic handwriting analysis, due to its noninvasive and readily accessible nature, has emerged as a vital adjunctive method for the early diagnosis of Parkinson's disease (PD). An essential step involves analysing subtle variations in signals to quantify PD dysgraphia. Although previous studies have explored extracting features from the overall signal, they may ignore the potential importance of local signal segments. In this study, we propose a lightweight network architecture to analyse dynamic handwriting signal segments of patients and present visual diagnostic results, providing an efficient diagnostic method.

To analyse subtle variations in handwriting, we investigate time-dependent patterns in local representation of handwriting signals. Specifically, we segment the handwriting signal into fixed-length sequential segments and design a compact one-dimensional (1D) hybrid network to extract discriminative temporal features for classifying each local segment. Finally, the category of the handwriting signal is fully diagnosed through a majority voting scheme.

The proposed method achieves impressive diagnostic performance on the new DraWritePD dataset (with an accuracy of 96.2%, sensitivity of 94.5% and specificity of 97.3%) and the well-established PaHaW dataset (with an accuracy of 90.7%, sensitivity of 94.3% and specificity of 87.5%). Moreover, the network architecture stands out for its excellent lightweight design, occupying a mere 0.084M parameters, with only 0.59M floating-point operations. It also exhibits nearly real-time CPU inference performance, with the inference time for a single handwriting signal ranging from 0.106 to 0.220 s.

We present a series of experiments with extensive analysis, which systematically demonstrate the effectiveness and efficiency of the proposed method in quantifying dysgraphia for a precise diagnosis of PD.