Robust Longitudinal Dementia Prediction under Systemic Missingness via Hierarchical Fusion and Test-Time Adaptation

Area of Science:

• Neuroscience
• Biomedical Informatics
• Machine Learning

Background:

• Longitudinal dementia progression prediction is crucial for clinical decisions.
• Existing models struggle with external data due to missing biomarkers and distribution shifts.
• Patient-specific variability further complicates accurate prediction.

Purpose of the Study:

• To develop a robust model for longitudinal dementia prediction that overcomes challenges of systemic missingness and patient variability.
• Introduce Progression-aware Feature Fusion with Test-Time Adaptation (ProFuse-TTA) for improved cross-dataset performance.

Main Methods:

• ProFuse-TTA utilizes a two-stage hierarchical Transformer architecture.
• Stage 1 learns temporal representations from irregular observations without imputation.
• Stage 2 fuses features using cross-feature attention with simulated modality dropout and employs test-time adaptation for per-individual calibration.

Main Results:

• ProFuse-TTA achieved superior cross-dataset performance in 8 out of 9 settings for predicting clinical diagnosis, MMSE, and hippocampal volume.
• The model ranked first in 14 out of 15 ablation scenarios, demonstrating robustness.
• Maintained high performance across varied input lengths and prediction horizons up to 6 years.

Conclusions:

• ProFuse-TTA offers a significant advancement in longitudinal dementia prediction, particularly in real-world scenarios with missing data.
• The proposed method enhances model generalizability and reliability for clinical applications.
• Availability of pretrained ADNI models facilitates further research and application.