ADNI Multimodal NCDE

Built a Neural CDE pipeline for next-visit ADAS-13 forecasting on irregular ADNI longitudinal data.

ADNI multimodal Neural CDE system architecture

Overview

This project models Alzheimer’s progression as a multimodal longitudinal prediction problem, combining clinical history with MRI-derived trajectory features on AWS-backed ADNI data.

My contribution

AWS data backbone, MRI trajectory feature pipeline, Neural CDE benchmarking, and ADNIMERGE MMSE ablation experiments.

Problem

ADNI data is irregular across visits and spread across modalities, which makes it difficult to model next-visit outcomes faithfully with simple fixed-step sequence assumptions.

Approach

Built an AWS S3/EC2 data backbone for ADNI MPRAGE sequences, PET scans, and clinical features (age, gender, genetics).
Extracted voxel-level MRI trajectories via flow matching and LoRA-finetuned point tracking before feeding irregular visit histories into a Neural CDE.
Benchmarked trajectory-only models against tabular baselines with ADNIMERGE MMSE features to isolate what drives forecasting gains.

Result

Trajectory-only Neural CDE models reached validation MAE of ~13–16 on next-visit ADAS-13 over 100 epochs. Adding ADNIMERGE tabular features (MMSE) cut error to ~7 MAE — the best-performing configuration in current benchmarking.

ADNI model training and validation metrics — Train MSE and validation MAE/RMSE over 100 epochs; tabular MMSE features cut Val MAE to ~7.

Stack

PyTorchtorchcdepandasAWS S3MRIADNI

Team repo (collaborative)