Generalizing Parkinson's disease detection using keystroke dynamics: a self-supervised approach

Area of Science:

• Neurology
• Machine Learning
• Digital Health

Background:

• Passive monitoring of touchscreen interactions offers a low-burden method for detecting neurological conditions like Parkinson's disease (PD).
• Current methods often require large, clinically labeled datasets from standardized environments, limiting scalability and generalizability.
• Self-supervised learning (SSL) presents a promising avenue to overcome data limitations in digital health applications.

Purpose of the Study:

• To validate a novel self-supervised learning method for analyzing touchscreen interactions.
• To assess the generalizability of the SSL approach across different datasets and subject groups.
• To reduce the dependency on extensive, clinically labeled datasets for neurological condition detection.

Main Methods:

• A new self-supervised loss function was developed, combining Barlow Twins loss and Dissimilarity loss.
• An encoder was pre-trained on unlabeled data from uncontrolled settings using the proposed SSL loss.
• The pre-trained model was then fine-tuned with clinically validated data and tested on independent datasets with controls and PD subjects.

Main Results:

• The proposed self-supervised learning approach demonstrated superior generalization capabilities compared to existing methods.
• Performance surpassed traditional supervised models, feature engineering strategies, and deep learning models pre-trained on Parkinsonian signs.
• The method proved effective in analyzing data from uncontrolled settings and across independent datasets.

Conclusions:

• The limitations of standardized data acquisition and labeled datasets hinder supervised model generalizability in neurological studies.
• Self-supervised models can learn robust patterns from data without requiring ground truth labels, enhancing applicability.
• This SSL approach can accelerate the clinical validation of touchscreen typing software for neurodegenerative diseases.