Abstract
Parkinson’s disease is the second most common neurodegenerative disorder in older adults. In addition to motor symptoms, it involves cognitive impairment, whose early detection is crucial for effective intervention. However, traditional assessment methods face significant limitations, such as evaluator subjectivity, interobserver variability, and lack of standardization, which hinder accurate and consistent diagnoses.This work presents the design and implementation of an AI-based system for automated segmentation and classification of the clock drawing task in the Montreal Cognitive Assessment (MoCA), aimed at supporting objective evaluation of cognitive decline. The solution integrates deep neural networks, using U-Net++ for segmenting components such as outline, numbers, and hands, achieving an average Jaccard Index (IoU) of 70.94%, improving up to 90% with pixel tolerances.
For cognitive classification, DenseNet121 with an Adapter7to3 module was employed. In binary classification (healthy control vs. Parkinson’s disease), the model achieved an AUC of 0.7009, with 74.6% sensitivity and 58.5% specificity. The multiclass classification (CN, MCI, PDD) reached a macro AUC of 0.8889, with metrics exceeding 75% across all categories.
Both models were integrated into a web platform that allows users to upload or draw images directly through the interface, providing automated results within seconds. The findings confirm the technical feasibility of the system and its potential as a complementary clinical tool. Future work includes dataset expansion, temporal stroke analysis, and longitudinal evaluation for monitoring cognitive progression.
| Date of Award | 23 May 2025 |
|---|---|
| Original language | Spanish |
| Awarding Institution |
|
| Supervisor | Antonio Martínez Torteya (Supervisor) |