Identifying dementia neuropathology using low-burden clinical data.

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Informatics

Background:

• Accurate identification of dementia neuropathology is crucial for developing effective treatments and conducting clinical trials.
• Current methods often require high-burden data, limiting their applicability in primary care settings.
• Semi-supervised learning models offer a promising approach to leverage low-burden data for improved generalizability.

Purpose of the Study:

• To develop and validate semi-supervised models for identifying dementia neuropathology using low-burden clinical data.
• To enhance the utility of data obtainable in primary care settings for dementia diagnosis.
• To improve the accuracy and generalizability of neuropathology prediction models.

Main Methods:

• Defined low-burden data as data reasonably obtainable in a primary care setting.
• Employed a semi-supervised learning paradigm, including clustering and prediction models.
• Trained models to identify and predict different neuropathology lesion types.

Main Results:

• A clustering model successfully identified two distinct patient groups: those enriched and those scarce in neuropathology.
• Semi-supervised prediction models demonstrated that low-burden data from multiple visits can predict neuropathology burden comparably to higher-burden data.
• The models achieved accurate predictions across various pathology types.

Conclusions:

• This research addresses a critical need by utilizing low-burden clinical data for neuropathology prediction, enhancing dementia screening.
• The findings support the use of semi-supervised learning for dementia neuropathology identification, aiding targeted therapies and clinical trials.
• Low-burden data, particularly longitudinal data, can provide accurate predictions of pathology load, with higher-burden data being most effective for vascular lesions.