IMPUTATION-FREE TRANSFORMER LEARNING ENABLES ROBUST ALZHEIMER'S DISEASE PREDICTION AND CALIBRATED UNCERTAINTY QUANTIFICATION ACROSS HETEROGENEOUS CLINICAL COHORTS

Clinical datasets for Alzheimer's disease are frequently characterized by missing values and high heterogeneity. Traditional approaches often rely on data imputation—filling in missing values—which can introduce systematic bias, distort relationships between clinical features, and produce overconfident, unreliable diagnostic predictions. This study addresses these challenges by proposing a novel architecture that operates directly on partially observed data.

The authors introduce NITROGEN, an imputation-free transformer model designed to handle multimodal clinical records. Unlike standard models, NITROGEN utilizes a dual-attention mechanism: masked attention to model dependencies within a single patient's features, and intersample attention to leverage relational structures across different patients. Furthermore, the model incorporates a modality-aware uncertainty adjustment, which scales the model's predictive uncertainty based on the importance of any missing data modalities. This ensures that when critical diagnostic information is absent, the model provides a more calibrated and cautious confidence score rather than a potentially misleading prediction.

NITROGEN was trained on the ADNI cohort and validated against the independent OASIS-3 and AIBL datasets. The results demonstrate that the model maintains competitive diagnostic and cognitive score prediction performance while significantly outperforming traditional tree-based ensemble methods in calibration and uncertainty quantification. The study also identifies key clinical predictors, such as cortical thickness in the temporal pole, age, and APOE genotype. The authors argue that for clinical deployment, researchers must prioritize calibration, interpretability, and cross-cohort reliability over raw accuracy metrics, as these factors are essential for building trust in AI-driven diagnostic tools.