Imputation-free transformer learning enables robust Alzheimer's disease prediction and calibrated uncertainty quantification across heterogeneous clinical cohorts

Core Problem: Data incompleteness and cohort heterogeneity undermine reliability of Alzheimer’s disease modeling

• Real-world clinical data suffer from pervasive missingness and heterogeneity, impeding trustworthy disease modeling and effective clinical evaluation;
• Conventional imputation strategies (e.g., mean/multiple imputation) introduce systematic bias, distort inter-feature correlations, and yield overconfident predictions—especially hazardous in diagnostic settings.

Methodological Innovation: NITROGEN enables end-to-end imputation-free multimodal learning

• Introducing NITROGEN: the first transformer architecture explicitly designed for partially observed clinical records without imputation;
• Key mechanisms:
• Masked attention: captures dynamic intra-patient dependencies among multimodal features (e.g., MRI, CSF, cognitive scales);
• Intersample attention: explicitly models inter-patient relational structure (e.g., shared progression patterns or biomarker profiles), mitigating cohort heterogeneity;
• Accepts raw, non-imputed, partially observed records—eliminating preprocessing-induced contamination.

Empirical Validation: Cross-cohort generalization and superior uncertainty calibration

• Training cohort: ADNI (N=7858 scans);
• Independent test cohorts: OASIS-3 (N=2675 scans) and AIBL (N=1286 scans);
• Tasks evaluated:
• Diagnostic classification (CN/MCI/AD);
• Continuous cognitive score prediction (e.g., MMSE, CDR-SB);
• Key results:
• Competitive discriminative performance vs. state-of-the-art imputation + tree ensembles (e.g., XGBoost): comparable AUC and MAE;
• Superior uncertainty quantification: coverage probabilities closer to nominal levels, lower quantile loss, and 37–52% reduction in Expected Calibration Error (ECE);
• Strong cross-cohort robustness—no catastrophic performance drop during out-of-distribution evaluation.

Significance and positioning

• Establishes a new paradigm for clinical AI: decoupling modeling from fragile imputation pipelines, enhancing trustworthiness and interpretability;
• Lies at the intersection of multimodal learning and uncertainty-aware deep learning;
• Published on arXiv:2607.11656v2 (replace-cross version), representing cutting-edge methodology research.