GNOVA model predicts Alzheimer's trajectories using only routine clinical data

A new AI framework called GNOVA (GRU-Neural ODE Variational Autoencoder) enables both reconstruction and forecasting of Alzheimer's disease progression using only routine clinical data—no expensive MRI, PET, or CSF biomarkers required. Trained on 1,727 patients from the ADNI dataset over a 10-year period, the model achieves mean absolute errors of 1.35 on the CDR-SB scale and 2.28 on the MMSE scale. The architecture combines a GRU encoder (handles irregular visit intervals and any number of inputs) with a Neural ODE decoder for continuous-time interpolation and extrapolation, all within a variational autoencoder for uncertainty quantification. Feature-ablation studies identified age, BMI, and APOE4 genotype as the strongest predictors, meaning the model can deliver clinically useful predictions in settings where advanced imaging is unavailable.

The work directly addresses two under-explored challenges in Alzheimer's modeling: bidirectional prediction (looking backward to fill gaps in patient history while also forecasting future decline) and calibration of prediction uncertainty. By not relying on costly modalities, GNOVA opens the door to deployment in low-resource clinics and developing regions. The continuous-time decoder lets clinicians query the model at any arbitrary time point, enabling personalized risk assessments and earlier interventions. The authors emphasize that their framework could be extended to other neurodegenerative diseases and electronic health record data, making scalable, AI-driven dementia care more accessible globally.