Abstract
Handwriting is significantly affected by neurological disorders (ND) such as Parkinson's disease (PD) and Alzheimer's disease (AD). Prior works have analyzed handwriting tasks using feature-based approaches or computer-vision techniques, but these methods have struggled to generalize across multiple datasets, particularly between temporal features represented as time-series and images. We propose a framework that leverages both time-series and images of handwriting through a joint classifier, based on a ResNet50 pretrained on ImageNet-1k. Binary classification experiments demonstrate state-of-the-art performances on existing time-series and image datasets, with significant improvement on specific drawing and writing tasks from the NeuroLogical Signals (NLS) dataset. In particular, the proposed model demonstrates improved performance on Draw Clock and Spiral tasks. Additionally, cross-dataset and multi-dataset experiments were consistently able to achieve high F1 scores, up to 98 for PD detection, highlighting the potential of the proposed model to generalize over different forms of handwriting signals, and enhance the detection of motor deficits in ND.
Abstract (translated)
手写受到神经性疾病（ND）如帕金森病（PD）和阿尔茨海默病（AD）的显著影响。先前的研究使用基于特征的方法或计算机视觉技术来分析手写任务，但这些方法在多个数据集之间推广时遇到了困难，尤其是在时间序列表示的时间特征与图像之间的推广方面存在挑战。我们提出了一种框架，该框架通过一个基于ImageNet-1k上预训练的ResNet50的联合分类器同时利用了手写的时序数据和图像信息。 二元分类实验表明，在现有的时间序列和图像数据集上取得了最先进的性能，并且在来自NeuroLogical Signals（NLS）数据集的具体绘制和书写任务中有了显著改进，特别是提出的模型在绘制钟表和螺旋线任务中的表现有所提升。此外，跨数据集和多数据集的实验始终能够达到较高的F1分数，在PD检测方面最高可达98分，这突显了所提出模型推广不同形式的手写信号以及增强ND运动功能障碍检测的能力。
URL
https://arxiv.org/abs/2509.16474
