Introduction
Neurodegenerative diseases (NDs), including Alzheimer’s Disease (AD) and Parkinson’s Disease (PD), affect millions of individuals globally. Early and accurate diagnosis is critical for intervention, yet current gold standards—such as neuroimaging and specialized clinical evaluations—are often expensive, invasive, or geographically inaccessible.
Handwriting is a complex task that integrates fine motor control with cognitive processing. While previous research has applied machine learning to handwriting analysis, many approaches rely on "fixed-size" snapshots that discard vital temporal information. This study addresses these limitations by proposing a "frame-level" analysis that treats handwriting signals as time-series data, enabling the detection of subtle motor impairments across varying signal lengths.
Methodology
The researchers utilized the NLS dataset from Johns Hopkins Hospital, comprising 113 participants across four categories: Healthy Controls (CTL), PD, AD, and Parkinson’s Disease Mimics (PDM). The participants performed 14 distinct tasks, including pointing, drawing spirals, and writing text.
The core of the approach involves treating raw handwriting data ($x, y$ coordinates and pen pressure) as multi-channel signals. These signals are decomposed into spectrograms using a Short-Time Fourier Transform (STFT). The model architecture compares a standard Convolutional Neural Network (CNN) with a CNN-BLSTM (Bidirectional Long Short-Term Memory) network to capture both spatial features and long-term temporal dependencies.
A key innovation is the use of a sliding window (ranging from 25ms to 1.5s) to create "frames" of spectrograms. This preserves the original signal's temporal resolution and allows the model to focus on fine-grained motor stability.
Key Results and Task Alignment
The study found that specific tasks are better suited for identifying different diseases. For instance, Point Tasks—which require holding a pen steady—were exceptionally effective for detecting Alzheimer’s and PDM, likely due to their ability to isolate fine motor tremors. In contrast, CopyReadText tasks, which involve simultaneous reading and writing, proved most effective for Parkinson’s detection by taxing both cognitive and motor systems.
The results indicate that the standard CNN generally outperformed the more complex CNN-BLSTM, suggesting that the spatial-temporal features within individual spectrogram frames are highly informative on their own. Optimal performance for AD detection reached an F1-score of 89.8% using Point Tasks, while PD was differentiated from Healthy Controls with a 74.5% F1-score on the CopyReadText task.
Discussion and Future Work
The research highlights the importance of velocity and pen pressure as diagnostic features. For AD, the combination of horizontal and vertical velocity with pressure was optimal. For PD, adding acceleration further improved model performance.
Despite these promising results, the study is limited by a relatively small sample size, particularly for the PDM group. Future research will focus on multimodal fusion—combining handwriting data with speech and eye-tracking signals—and developing "Explainable AI" tools to help clinicians interpret model outputs. This work paves the way for accessible, non-invasive screening tools for neurodegenerative conditions.
Summary updated