Alzheimer detection system

detection
home

Inspiration

Late-stage Alzheimer's diagnosis is too late. Current single-test approaches (MRI-only or cognitive tests alone) miss the disease's multimodal complexity. Why Mixture-of-Experts? Neurologists don't average all evidence equally, they prioritise strong signals. When MRI shows severe atrophy, they weight it heavily. When handwriting reveals tremor, they trust that signal. MoE learns this adaptive reasoning: a gating network that asks "Which expert should I trust for THIS patient?"

What it does

Analyzes three biomarkers in <10 seconds: MRI scans - Structural brain atrophy Handwriting - 450-dim motor control features Clinical scores - MMSE, CDR, age, memory Results: 63.1% accuracy, beats single-modality by 5.3%

How we built it

Architecture: Sparse MoE with 3 experts (ResNet-18 for MRI, MLPs for handwriting/clinical) + gating network that learns dynamic routing Key Techniques: Temperature scaling (T=3.0) fixes overconfidence Feature normalization (handwriting was exploding to 2.7M) Top-k=2 sparse routing for efficiency Stack: PyTorch, Flet UI, 50 epochs, AdamW optimizer

Challenges we ran into

1.Exploding features - Handwriting values hit 2.7M → Fixed with z-score normalization

100% overconfidence - Dangerous in healthcare → Temperature scaling
Finding speech dataset that is open source ## Accomplishments that we're proud of
63.1% accuracy with interpretable expert routing
27% clinical threshold validated on real patients
Safety-first: calibrated confidence, no dangerous overconfidence
Real-time: <10 seconds on CPU ## What we learned
MoE models really do better than simple ensembles: letting the system route information dynamically gave us about 1.9% better performance than just averaging everything.
Good normalization beats adding more layers: fixing the way features are scaled made a much bigger difference (about 10×) than stacking extra layers.
Calibration matters a lot: it’s far more useful to have a model that’s 60% confident and actually right than one that claims to be 100% confident but is wrong.