An AI system can now flag Alzheimer's disease by pinpointing the exact moments in a handwriting sample that look abnormal, instead of grading the whole page.
Researchers built a model called NormPaST-Risk that analyzes online handwriting captured by a digitizing tablet, which records both pen strokes on paper and the pen's movement in the air between strokes. A multi-scale encoder reads stroke dynamics at several time resolutions, while a second component separates on-paper motor control from in-air planning and transitions. The model learns what normal handwriting dynamics look like from healthy participants, then flags segments that deviate from that baseline using an expert module trained without manual labeling of which segments matter. On the DARWIN benchmark, a dataset used to test Alzheimer's classification from handwriting, it outperformed existing detection methods.
The interesting part isn't the accuracy gain, it's the interpretability. Most handwriting-based dementia detectors output one score for an entire writing sample, which is hard to trust and impossible to double-check. This one can project the flagged strokes and pauses back onto the original trajectory, giving a clinician actual evidence instead of a black box verdict.
That's still a single-benchmark result on one dataset, a long way from a tool any doctor's office would trust to catch Alzheimer's.