NITROGEN: Imputation-free transformer predicts Alzheimer's with calibrated uncertainty

Alzheimer's disease prediction from real-world clinical data is plagued by missing values and cohort heterogeneity. Traditional imputation methods introduce systematic bias and overconfident predictions. To address this, researchers from Switzerland and Vietnam developed NITROGEN, an imputation-free transformer that leverages masked and intersample attention to learn from partially observed multimodal records without filling in missing data. It jointly models within-patient feature dependencies and between-patient relational structure, enabling robust diagnostic classification and cognitive score prediction. The model was trained on 7,858 scans from the ADNI dataset and validated on two independent cohorts: OASIS-3 (2,675 scans) and AIBL (1,286 scans).

NITROGEN demonstrated superior calibration and uncertainty quantification compared to tree-based ensemble methods, while maintaining competitive discriminative performance across cohorts. Cross-cohort and cross-method analyses revealed cortical thickness in the temporal pole, age, and APOE genotype as important but individually insufficient features for Alzheimer's classification. The team also introduced a modality-aware uncertainty adjustment that increases predictive uncertainty proportionally to the importance of absent modalities, enabling calibrated confidence even when key diagnostic information is missing. The study emphasizes that clinical deployment of AI models should evaluate calibration, interpretability, and cross-cohort reliability—not just accuracy—to ensure trustworthy predictions in heterogeneous real-world settings.