DKT

DKT transfers multimodal biomarker information between related neurodegenerative diseases to infer multimodal biomarker trajectories and predict unseen biomarkers in rare conditions such as Posterior Cortical Atrophy (PCA) using data from common disorders like typical Alzheimer's disease (tAD).

Key Features:

• Joint-disease generative model: leverages shared biomarker relationships across diseases to enable cross-disease inference.
• Cross-disease training on combined datasets: trains on combined multimodal tAD data from the TADPOLE Challenge and unimodal PCA data from the Dementia Research Centre (DRC).
• Multimodal trajectory estimation from unimodal data: estimates plausible multimodal biomarker trajectories in diseases with only unimodal measurements, such as MRI-only PCA cohorts.
• Validation with simulations and patient cohorts: validated using synthetic data simulations and real patient datasets including TADPOLE and PCA cohorts.
• Prediction of unseen biomarkers: infers biomarkers not observed in the smaller target-disease dataset.
• Parameter recovery in simulations: accurately estimates ground truth model parameters in simulated environments.
• Generalizability: methodology applicable to related neurodegenerative diseases beyond Alzheimer's disease variants.

Scientific Applications:

• Estimating disease progression in rare disorders: infers biomarker trajectories for conditions such as PCA by transferring information from tAD.
• Augmenting clinical cohorts with predicted biomarkers: provides inferred multimodal biomarker profiles when only unimodal data are available.
• Supporting translational research and personalized medicine: informs development of targeted therapeutic strategies and personalized approaches by improving biomarker-based characterization.

Methodology:

DKT employs a joint-disease generative model trained on combined multimodal tAD data from the TADPOLE Challenge and unimodal PCA data from the Dementia Research Centre (DRC), with evaluation via synthetic data simulations and real TADPOLE and PCA patient cohorts to estimate multimodal biomarker trajectories and predict unseen biomarkers.