Researchers have adapted CLIP, the image-recognition AI behind many text-to-image tools, to screen for Alzheimer's disease from handwriting samples.
The team built a lightweight add-on called a Cross-Layer Fusion Adapter, which plugs into CLIP's frozen visual encoder and adds depthwise convolution layers to catch both small stroke irregularities and bigger-picture handwriting structure. They tested it on the Darwin dataset, a handwriting-based dementia screening dataset, using a subject-disjoint cross-task setup: the model trained on one handwriting task, like copying a sentence, and got evaluated on a completely different task, like drawing a clock, for people it had never seen during training. Averaged across all 600 possible task pairings, the adapted model reached 74.63% AUC, 74.85% accuracy, and 73.72% F1 score, beating the best competing approach by about 2 percentage points on each measure.
Most handwriting-based Alzheimer's screening tools are trained and tested on the same task, which is a problem for real clinics that cannot guarantee a patient will do the exact writing exercise a model expects. Showing that a general-purpose vision-language model can transfer across handwriting tasks, without retraining for each one, is a step toward screening tools flexible enough for actual clinical use.
Still, a 74% AUC is a research benchmark, not a diagnosis - this is early-stage evidence that repurposing off-the-shelf AI models might work for handwriting analysis, not proof that it's ready for a doctor's office.