Cross-dataset Evaluation of Dementia Longitudinal Progression Prediction Models

Area of Science:

• Neuroscience
• Biomedical Informatics
• Machine Learning

Background:

• Accurate prediction of Alzheimer's Disease (AD) progression is vital for timely therapeutic interventions.
• The TADPOLE challenge benchmarked 92 algorithms using multimodal biomarkers for predicting clinical diagnosis, cognition, and ventricular volume.
• The winning FROG algorithm employed a Longitudinal-to-Cross-sectional (L2C) transformation, differing from methods fitting entire longitudinal histories.

Purpose of the Study:

• To evaluate the generalizability of the FROG algorithm and its variants on external datasets.
• To introduce and assess a novel L2C feedforward neural network (L2C-FNN) variant.
• To compare the predictive performance of L2C-FNN against established methods for AD progression prediction.

Main Methods:

• The FROG algorithm's L2C transformation was applied to convert longitudinal patient data into fixed-length feature vectors.
• A L2C feedforward neural network (L2C-FNN) was developed, integrating XGBoost models with a feedforward network.
• The L2C-FNN and AD-Map models were trained on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset and validated on three independent external datasets.

Main Results:

• L2C-FNN demonstrated superior performance in predicting clinical diagnosis across external datasets.
• Both L2C-FNN and AD-Map achieved top performance in predicting cognition and ventricular volume.
• L2C-FNN maintained strong predictive accuracy irrespective of the number of observed timepoints and for long-term predictions (0-6 years).

Conclusions:

• The L2C-FNN model exhibits excellent generalizability and robust performance for long-term Alzheimer's Disease progression prediction.
• This approach offers a promising tool for early intervention strategies by accurately forecasting disease trajectory.
• Publicly available pretrained ADNI models facilitate further research and clinical application.