Micrographia in Parkinson's Disease: Automatic Recognition through Artificial Intelligence

Francesco Asci, Gaetano Saurio, Giulia Pinola, Marco Falletti, Alessandro Zampogna, Martina Patera, Francesco Fattapposta, Simone Scardapane, Antonio SuppaMovement 2025 · DOI

TL;DR

This paper introduces an AI-powered, pen-and-paper handwriting analysis tool to automatically detect and quantify micrographia in Parkinson's Disease (PD). The AI model demonstrates early detection of micrographia, its progression with disease severity, and its limited response to L-Dopa, offering a promising, low-cost solution for PD assessment and telemedicine.

What Problem Does This Paper Address?

Parkinson's Disease (PD) is often accompanied by micrographia, a characteristic shrinking of handwriting. However, there's ongoing debate about whether micrographia appears early in the disease, how it progresses, and its responsiveness to L-Dopa treatment. Traditional methods for assessing micrographia are often subjective and prone to inter-rater variability. This study aims to address these gaps by developing and validating a non-invasive, cost-effective Artificial Intelligence (AI)-based handwriting analysis tool to objectively detect and quantify micrographia in PD patients. This could pave the way for more accessible and objective diagnostic and monitoring tools, especially for telemedicine applications.

Methods

The researchers employed a Convolutional Neural Network (CNN) for their AI-based handwriting analysis. The process involved:

  • Data Acquisition: Participants completed two standardized pen-and-paper handwriting tasks.
  • Digitalization: Handwriting samples were digitized using smartphone cameras.
  • Preprocessing: Samples underwent cleaning, grey-binarization, and cropping to isolate individual handwriting instances.
  • AI Model Training: A CNN model, specifically EfficientNet as the backbone, was fine-tuned on the collected handwriting data. The model was trained to classify handwriting samples into categories such as healthy subjects (HS) vs. PD patients, early-stage PD vs. mid-advanced stage PD, and patients in OFF vs. ON L-Dopa states.
  • Evaluation: The model's performance was assessed using Receiver Operating Characteristic (ROC) curves, calculating metrics like sensitivity, specificity, accuracy, and area under the curve (AUC).

For comparison, a clinically-based perceptual analysis was also performed by a neurologist using a modified visual analogue scale (VAS) to evaluate handwriting features like average height, length, and progressive size variation.

Datasets

The study utilized a dataset of handwriting samples collected from:

  • 57 Parkinson's Disease (PD) patients:
    • 30 early-stage patients (Hoehn & Yahr stage ≤ 2)
    • 27 mid-advanced stage patients (Hoehn & Yahr stage > 2)
  • 25 age- and sex-matched healthy controls (HS)

Participants were chronic L-Dopa users and were studied in both OFF (at least 12 hours after last dose) and ON (1-2 hours after first dose) states. The paper does not introduce a new benchmark dataset but uses existing pen-and-paper tasks adapted for an ecological scenario.

Key Insights and Ideas

  • AI as an Objective Tool: The study strongly advocates for AI-based handwriting analysis as a superior, objective method compared to traditional perceptual assessments, especially for detecting subtle changes in PD.
  • Early Detection of Micrographia: The AI model was able to detect micrographia in PD patients, suggesting it manifests early in the disease, potentially even before clinical diagnosis.
  • Disease Progression Matters: The AI model could differentiate between early and mid-advanced stages of PD based on handwriting, indicating that micrographia worsens as the disease progresses. This is a significant finding that traditional methods struggled to capture.
  • Limited L-Dopa Response: Both clinical and AI-based analyses indicated that L-Dopa treatment had a limited effect on micrographia, particularly on progressive micrographia. This challenges some previous assumptions about the drug's impact on this specific symptom.
  • Ecological Validity: The use of simple pen-and-paper tasks in an "ecological scenario" (participants' homes) makes the approach highly practical and suitable for telemedicine.

Key Results

  • PD vs. Controls: The AI model achieved high accuracy (91%) in distinguishing PD patients from healthy controls.
  • Disease Stage Differentiation: The AI model showed moderate accuracy (77%) in differentiating between early-stage and mid-advanced stage PD patients.
  • L-Dopa State: The AI model was not able to reliably detect differences between patients in the OFF and ON L-Dopa states, suggesting limited responsiveness of micrographia to this treatment.
  • Clinical vs. AI: While both methods confirmed micrographia in PD, the AI approach provided more granular insights into its early onset and progression. Clinically-based analysis showed comparable VAS scores between early and mid-advanced stages, which the AI analysis contradicted.

Limitations and Future Work

  • Sample Size: The authors acknowledge that the number of patients included in the cohorts might be considered small, and larger studies could potentially reveal finer distinctions regarding disease progression and L-Dopa response.
  • Subjectivity in Clinical Assessment: The paper notes that inherent weaknesses in perceptual clinical assessment might have contributed to some of the findings, particularly the lack of detected L-Dopa effect.
  • Scope of Handwriting Abnormalities: The study focused primarily on micrographia. The authors suggest that the term "PD dysgraphia" might be more appropriate to encompass the broader spectrum of writing disturbances in PD, including dynamic and kinematic features.
  • Pathophysiology: While the study provides evidence for micrographia's role in PD, it does not definitively pinpoint specific brain structures responsible for its pathophysiology.
  • Future Directions: The authors suggest exploring different AI algorithm setups, generating more handwriting samples, and developing revised AI kinematic tools. They also propose using sequential AI models trained on recurrent handwriting data collected over time to track disease progression more effectively. The potential for using this AI tool for telemedicine is highlighted as a key future application.

Summary updated

Abstract

AI-based handwriting analysis is a valuable non-invasive and cheap tool for detecting and quantifying micrographia in PD, for telemedicine purposes.
BibTeX
@inproceedings{asci2025micrographia,
  author = {Francesco Asci and Gaetano Saurio and Giulia Pinola and Marco Falletti and Alessandro Zampogna and Martina Patera and Francesco Fattapposta and Simone Scardapane and Antonio Suppa},
  title = {{Micrographia in Parkinson's Disease: Automatic Recognition through Artificial Intelligence}},
  booktitle = {Movement},
  year = {2025},
  doi = {10.1002/mdc3.70208},
}
Figure