Caractérisation de la Maladie d'Alzheimer à travers l'Ecriture manuscrite acquise sur Tablette graphique

Christian Kahindo, Sonia Garcia-Salicetti, Mounîm A. El‐Yacoubi, Victoria Cristacho-Lacroix, Hélène Kerhervé, Anne‐Sophie RigaudHAL 2017

TL;DR

This paper explores using handwriting analysis from digital tablets to detect early-stage Alzheimer's disease (AD) and Mild Cognitive Impairment (MCI). By analyzing spatio-temporal writing parameters, the researchers developed an automated method to identify distinct cognitive profiles, offering a potential non-invasive diagnostic tool.

What Problem Does This Paper Address?

Diagnosing Alzheimer's disease (AD) and Mild Cognitive Impairment (MCI) in their early stages is a significant challenge. The insidious onset of AD means that by the time symptoms are obvious, the disease has often progressed for years. Current diagnostic methods can be invasive or require specialized equipment. Handwriting, a complex motor skill, is known to be affected by neurological conditions. However, the literature on using fine-grained handwriting analysis from digital tablets to differentiate between AD, MCI, and healthy controls is relatively sparse. This paper aims to bridge this gap by characterizing handwriting patterns associated with early AD and MCI, leveraging the rich spatio-temporal data captured by graphic tablets.

Methods

The core of the methodology involves extracting a comprehensive set of spatio-temporal parameters from handwriting samples acquired on a graphic tablet. These parameters fall into three categories:

  • Kinematic: These include velocity (in x, y, and module), acceleration, and "jerk" (the rate of change of acceleration) in both x and y directions. These are calculated from the pen's position over time.
  • Spatial: These encompass direction, curvature, length of intra-word and inter-word spaces, and the complexity of the pen's trajectory when the pen is lifted ("pen-up").
  • Dynamic: This category includes pressure exerted on the tablet and variations in pressure.

A novel aspect of their approach is the automatic quantification of these parameters. Instead of relying on simple averages and standard deviations, they use a k-means clustering algorithm to group the local values of each parameter into 5 "bins." This creates a compact, 5-dimensional representation of the parameter's distribution for each participant.

For classification, they employ Hierarchical Agglomerative Clustering (CAH) with Ward's criterion. To select the most discriminative parameters, they iteratively add parameters that maximize the normalized mutual information (IMN) between the clusters and the cognitive classes (AD, MCI, Healthy Control). This iterative selection process aims to find the optimal combination of parameters that best separates the cognitive groups.

Datasets

The study utilized data from 182 participants. The breakdown of participants included:

  • 29 patients with early-stage Alzheimer's disease (AD)
  • 86 individuals with Mild Cognitive Impairment (MCI) (further categorized into executive, amnestic, and multi-domain MCI)
  • 27 healthy controls (HCC)
  • Other participants with Parkinson's disease, combined AD and Parkinson's, or unknown diagnoses were excluded from the primary analysis.

The data was collected at Hôpital Broca. The paper also mentions that participants will be followed up longitudinally at 12 months, suggesting future work will utilize this longitudinal data.

Key Insights and Ideas

  • Handwriting as a sensitive biomarker: The study reinforces the idea that subtle changes in handwriting, particularly its dynamic and spatio-temporal aspects, can reflect underlying neurological changes associated with AD and MCI.
  • "Pen-up" movements are informative: Parameters related to the pen's movement when it's not in contact with the tablet (e.g., jerk in air, number of strokes) are shown to be significant discriminators, suggesting that even the "idle" phases of writing contain valuable diagnostic information.
  • Jerk is a key indicator of fluidity: The "jerk" parameter, representing the variation in acceleration, consistently emerged as a strong indicator of writing fluidity and was found to be important in differentiating cognitive groups. This aligns with the understanding that AD affects fine motor control.
  • MCI is not a monolithic group: The analysis revealed that MCI patients could be clustered into two distinct groups: one more similar to AD patients and another closer to healthy controls, highlighting the heterogeneity within MCI.
  • Automated parameter selection is crucial: The use of normalized mutual information for iterative parameter selection is a key methodological contribution, allowing for the identification of the most informative features without manual feature engineering.

Key Results

The study identified several key parameters that effectively differentiated cognitive groups:

  • For AD vs. Healthy Controls (HCC):
    • Pen-up trajectory: Vertical jerk, normalized jerk (by length and duration), and the number of strokes (defined by periods of zero vertical velocity in air).
  • Tablet trajectory: Module of velocity.
  • For AD vs. MCI vs. HCC:
    • Pen-up trajectory: Number of points with zero vertical acceleration and inter-word time.
  • Tablet trajectory: Vertical jerk.

The clustering results showed:

  • When comparing HCC and AD, Cluster 1 predominantly contained HCC (24/28), and Cluster 2 predominantly contained AD (23/29).
  • When including MCI, Cluster 2 contained most AD patients (21/29) and a significant portion of MCI patients (38/87), while Cluster 3 contained most HCC (24/28) and some MCI patients (47/87). This confirms the two distinct MCI profiles.

Limitations and Future Work

  • Cross-sectional study: The current study is cross-sectional. The authors acknowledge the need for a longitudinal study to determine the predictive value of these parameters for disease progression.
  • Parameter selection: While the method is automated, the paper notes that the "best" combination of parameters was identified for specific comparisons (AD vs. HCC, and AD vs. MCI vs. HCC). Further exploration of parameter combinations for other comparisons might be beneficial.
  • Generalizability: The study was conducted with a specific cohort at Hôpital Broca. Future work should aim to validate these findings on larger, more diverse datasets.
  • Exploring other tasks: The current analysis primarily focused on copying a text. The authors plan to investigate other handwriting tasks (free writing, loops, circles, spirals) to extract additional discriminative parameters.
  • Predictive value: The ultimate goal is to determine which parameters have predictive value for the onset and progression of Alzheimer's disease, which will be addressed in the planned longitudinal study.

Summary updated

Abstract

International audience
BibTeX
@inproceedings{kahindo2017caracterisation,
  author = {Christian Kahindo and Sonia Garcia-Salicetti and Mounîm A. El‐Yacoubi and Victoria Cristacho-Lacroix and Hélène Kerhervé and Anne‐Sophie Rigaud},
  title = {{Caractérisation de la Maladie d'Alzheimer à travers l'Ecriture manuscrite acquise sur Tablette graphique}},
  booktitle = {HAL},
  year = {2017},
}
Figure