NeuroFlow

This week NeuroFlow finally went from just being an idea to something actually real. I got it set up with the TADPOLE Alzheimer’s data and made sure the model only sees past patient visits when predicting future ones, so there’s no accidental cheating by peeking at data it shouldn’t have. I also cleaned up the Git repo, locked down the raw medical data with .gitignore, and got a safe first commit in. My biggest realization this week was that the simple baseline models are honestly really strong for Alzheimer’s prediction, since a lot of the biomarkers change pretty slowly over time. So instead of forcing some fancy deep model to beat them head-on, I flipped the whole approach. Now the deep model just learns the gap between the baseline prediction and the real future outcome, which is way smarter than brute-forcing it.I wrapped up Phase 1 by creating the baseline predictions and residual targets, then built Phase 2: a residual time-aware Transformer that looks at past visits, time gaps, and baseline predictions all together. Even in a tiny 300-patient test run it already improved ADAS13, hippocampus, ventricles, and slightly whole-brain volume, which felt really good. Next up I want to push it on a bigger run and then compare it against a residual Neural CDE. Wish me luckk!