University of the Aegean
Revealing common mechanisms between neurodegenerative diseases and artistic behaviors using fuzzy knowledge representation
Abstract
dc:descriptionNeurodegenerative diseases (NDs), including Parkinson's Disease (PD), Alzheimer's Disease (AD), Frontotemporal Dementia (FTD), Corticobasal Degeneration (CBD), and Lewy Body Dementia (LBD), represent a diverse class of progressive brain pathologies. Traditionally, clinical approaches to NDs have had limited power in detecting overlaps, especially considering the variety of heterogeneous sources of data in the areas of behavior, cognition, and neurology. At the same time, the neurological deficits leading to NDs impact creative expression through lesions of brain regions that control motor processes. Uncovering the overlap between the pathology-affected regions and the regions that control motor function, which is involved in artistic expression, can be beneficial in terms of early biomarker discovery as well as the identification of correlations between the disease and creative behaviors. In the present research, the overlap in neural bases of neurodegenerative diseases and artistic behaviors will be studied by means of fuzzy knowledge representation. The current study targets two specific research aims: (a) building models based on fuzzy ontology for the representation of biological, clinical, and artistic knowledge concerning neurodegeneration; and (b) developing a reliable and clinically actionable framework for the diagnostic testing of PD using artistic behaviors represented by hand drawings.For addressing the stated goals, three AI methods were used to build classification models for PD diagnostics on datasets with increasing complexity: supervised machine learning algorithms (ML), large language models (LLMs), and fuzzy ontology-based reasoning. For the first dataset, statistical features were obtained by measuring kinematic characteristics of hand-drawn spirals. In the second experiment, waves drawn on digitized paper were analyzed. In the third and final experiment, stylus drawing was considered, where both kinematic and geometrical characteristics were measured. To address the problem of low availability of training samples, synthetic data was augmented using Conditional Tabular GAN (CTGAN). High F1-scores (0.70–0.94) for machine learning (ML) classifiers trained on synthetic datasets and applied to real-world data indicate high generalizability. Ontological models, built using the fuzzyOWL2 plugin in Protégé, used rule-based reasoning on knowledge represented by triangular membership functions. Superior interpretability and high F1-score of 0.91 on the most complex dataset prove high robustness of the models on uncertain input data. For zero-shot generation, the large language model ChatGPT was tested on drawing features using natural language input prompts. Accuracy achieved by zero-shot models was modest (up to 0.77 F1-score). Lack of domain-specific knowledge made these models suboptimal for PD diagnostics.Comparison of the experiments reveals consistent patterns. According to SHAP analysis, features with highest discriminatory power included signal entropy, smoothness, and kurtosis, corroborating the ontology-based rules and qualitative findings. Fuzzy ontologies offered an advantage of interpretability in terms of rules defined by domain experts. Ontologies also outperformed the other models on highly varying stylus data. As opposed to crisp ontologies and ML approaches, fuzziness of membership functions facilitated smooth decision-making in the presence of noisy data. The obtained results highlight the potential of knowledge-intensive approaches in clinical AI.Biomarkers identified during the research, namely spiral irregularity, wave energy, and pressure dynamics, relate to processes controlled by the basal ganglia circuits. The ontology developed as part of this thesis opens the opportunity for future research combining neuroimaging data and drawing-based performance metrics to study the relationships between structural brain differences in artists and pathological brain degeneration in neurodegenerative diseases. However, limitations include small sizes of real-world datasets, lack of adaptiveness in ontology and zero-shot LLM prompt engineering, and no domain-specific tuning of LLMs. To solve these issues, larger, longitudinal datasets including motor performance in a variety of activities should be collected, while multiple modalities like voice and gait measurements should be incorporated. Following the broader research agenda of identifying the overlaps between neurodegenerative diseases and creative behaviors, a future direction includes building a neuro-symbolic architecture combining language-based reasoning of LLMs with fuzzy ontology.
Degree
thesis:*- Grantor dc:publisher
- University of the Aegean
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Koletis, Adam
- Κωλέτης, Αδάμ
Subjects
dc:subject × 26- Νευροεκφυλιστικές Ασθένειες (NDs)
- Νόσος του Πάρκινσον (PD)
- Καλλιτεχνική Συμπεριφορά, Διάγνωση Βασισμένη σε Σχέδια
- Ασαφής Αναπαράσταση Γνώσης
- Μεγάλα γλωσσικά μοντέλα (LLMs)
- Ερμηνεύσιμη Τεχνητή Νοημοσύνη (XAI)
- Νευρο-Συμβολικά Συστήματα
- Γράφοι γνώσης
- Ψηφιακοί βιοδείκτες
- Neurodegenerative Diseases (NDs)
- Parkinson's disease (PD)
- Artistic Behavior; Drawing-Based Diagnosis
- Fuzzy Knowledge Representation
- Fuzzy Ontologies
- Machine learning (ML)
- Large Language Models (LLMs)
- Explainable AI (XAI)
- Neuro-Symbolic Systems
- Knowledge graphs
- Digital biomarkers
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62054