Each patient is modelled as a heterogeneous graph — biomarker, speech, gait and demographic nodes connected by learned weighted edges — capturing inter-modal correlations that single-modality models miss.
pTau181 and NfL — two numeric inputs drawn from a routine blood draw.
A 60-second Cookie Theft picture description, reduced to acoustic and linguistic features.
30 seconds of smartphone accelerometry, reduced to spatiotemporal gait features.
Age, sex, APOE4 genotype and MMSE score.
Retrospective research results on held-out research datasets. They are not clinical performance claims, and NFAD is not available for clinical use.
Alzheimer's pathology begins 15–20 years before symptoms, yet 90% of patients are diagnosed at a moderate or late stage. NFAD is being researched as a possible primary-care triage aid, where a higher-risk estimate could support earlier specialist referral.
Pathological damage begins long before the first symptom appears.
Most patients are identified only at a moderate or late stage.
The cost of confirmatory amyloid PET, inaccessible to most patients.
Average specialist wait times for those seeking assessment.
NFAD was developed in sequence, with a mid-project fix to eliminate data leakage before evaluation on two held-out research test sets. Final validation on DementiaBank is on hold pending incorporation.
Graph design and the multimodal data pipeline established.
Initial model trained and tuned on ADNI data.
Data leakage resolved by reducing model capacity and rebuilding the train–test split protocol.
Performance measured on held-out ADNI and Bio-Hermes-001 research test sets. Not a clinical performance claim.