Cross-Lingual Parkinson’s Disease Classification using Few-Shot Transfer Learning with Interpretable and Non-Interpretable Features
Keywords
Loading...
Authors
Issue Date
2025-02-01
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Parkinson’s Disease (PD) is the second-most prevalent neurodegenerative dis
order worldwide, affecting approximately 1% of individuals over the age of 60.
Speech biomarkers are among the earliest signs of PD, allowing for voice di
agnostics for early detection to start a timely intervention. However, linguistic
differences in PD symptoms complicate classification in new languages. Transfer
learning (TL) offers a promising approach for cross-lingual classification, though
it has not yet been extensively explored in the context of PD classification. Re
cent advancements in Artificial Intelligence (AI) in healthcare have further en
couraged its adoption in clinical practice. This thesis investigates the potential
of few-shot TL for PD classification across languages using speech data, focus
ing on the performance of interpretable (IFM) and non-interpretable (NIFM)
feature models. Through a series of experiments, the impact of increasing the
f
ine-tuning set size on classification performance was explored, with the model
trained on a base dataset and fine-tuned incrementally on a target dataset in
a different language. This study demonstrates the feasibility of cross-lingual
few-shot learning for PD classification, with NIFM showing slight but consis
tent advantages. The findings reveal that the zero-shot scenario, in which no
target language fine-tuning occurs, results in chance-level classification perfor
mance for both IFM and NIFM. However, when fine-tuning using data from the
target set, both feature models show significant performance gains. This per
formance improvement flattens as the fine-tuning set exceeds 50% of the target
data. There is a consistent, though modest, advantage of NIFM over IFM across
all experiments. Both feature types demonstrate improvement with fine-tuning,
with NIFM maintaining a slight advantage even in the zero-shot setting. Fur
thermore, the results indicate that the relative effectiveness of IFM and NIFM
is highly dataset-dependent. While base set performance declined as fine-tuning
progressed, incorporating a balanced mix of base and target samples during fine
tuning partially mitigated this effect. Visual analyses and feature importance
graphs highlighted language-dependent differences in the way PD symptoms
manifest in voice features, demonstrating the challenge of cross-lingual diagno
sis. This study emphasized the importance of data quality for effective TL,
highlighting the need for dataset developers to establish and follow standardized
data collection protocols. Future work will focus on developing more generaliz
able feature representations and exploring advanced fine-tuning approaches for
cross-lingual voice-based PD classification.
Keywords— Parkinson’s disease, speech classification, cross-lingual, few-shot transfer
learning, interpretable features, non-interpretable features
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen