Develop a diagnostic tool for dementia using machine learning and non-imaging features.

Area of Science:

• Gerontology
• Medical Informatics
• Machine Learning

Background:

• Early identification of Alzheimer's disease and mild cognitive impairment is crucial for timely interventions and cost reduction.
• Current neuroimaging diagnostic tools face challenges in cost, accessibility, and clinical integration.
• A novel approach is needed to overcome the limitations of existing diagnostic methods.

Purpose of the Study:

• To develop and evaluate a machine learning-based diagnostic tool for early dementia detection.
• To utilize non-imaging factors for a more accessible and cost-effective diagnostic solution.
• To implement an online tool with a user-friendly interface for widespread clinical use.

Main Methods:

• Trained and validated multiple machine learning models (logistic regression, SVM, neural network, random forest, XGBoost, LASSO, best subset) on a dataset of 654 elderly participants.
• Utilized 70 demographic, cognitive, socioeconomic, and clinical features.
• Externally validated models on a separate dataset of 1,100 participants and identified key predictive features using LASSO and best subset models.

Main Results:

• The Support Vector Machine (SVM) with a polynomial kernel demonstrated superior performance (AUROC = 0.9213) across datasets.
• Neural network (AUROC = 0.9435) and XGBoost (AUROC = 0.9398) also showed strong predictive power on the nursing home dataset.
• Models like SVM, neural network, and random forest proved robust on an unbalanced community dataset, with SVM (polynomial kernel) being the overall best performer.

Conclusions:

• The developed non-imaging-based diagnostic tool effectively predicts dementia outcomes.
• The tool's online implementation offers zero-barrier access, enhancing dementia diagnosis and care quality.
• This innovative approach promises to reduce healthcare costs associated with dementia.