Enhancing identification performance of cognitive impairment high-risk based on a semi-supervised learning method

Area of Science:

• Computational neuroscience
• Machine learning in healthcare
• Cognitive impairment diagnostics

Background:

• Cognitive assessment is crucial for early detection of cognitive impairment, including Alzheimer's and Lewy body dementia.
• Current large-scale screening relies on cognitive scales, which can be insensitive or costly.
• Machine learning applications for cognitive function assessment are underexplored in screening, often needing expert annotations.

Purpose of the Study:

• Introduce a semi-supervised learning algorithm, pseudo-label with putback (SS-PP), to enhance model efficiency.
• Improve prediction of high risk of cognitive impairment (HR-CI) by leveraging unlabeled data.
• Explore cost-effective strategies for cognitive healthcare domains.

Main Methods:

• Developed and evaluated a semi-supervised classification algorithm (SS-PP) using 189 labeled and 215,078 unlabeled real-world samples.
• Compared SS-PP performance against 14 traditional supervised machine-learning methods and other advanced semi-supervised algorithms.
• Utilized physical examination data within the prediction model.

Main Results:

• The optimal SS-PP model, based on Gradient Boosting Decision Tree (GBDT), achieved an Area Under the Curve (AUC) of 0.947.
• Demonstrated an average AUC improvement of 8% compared to supervised learning models.
• Achieved state-of-the-art performance among evaluated semi-supervised methods.

Conclusions:

• Pioneered the use of limited labeled data for HR-CI predictions through semi-supervised learning.
• Highlighted the benefits of incorporating physical examination data for improved cognitive impairment risk prediction.
• Implications for developing cost-effective screening and management strategies for cognitive diseases.