This paper presents a patient-level study on first-level screening for Alzheimer’s disease from offline handwriting images. The analysis is based on DARWIN-I handwriting samples and compares two Ultralytics YOLO-based image classifiers, YOLO11 and YOLO26, under a strict patient-level split designed to prevent identity confounding. The dataset is partitioned into 140 training subjects, 17 validation subjects, and 17 test subjects, corresponding to 1744, 208, and 221 images, respectively. Two experimental scenarios are evaluated. In the baseline setting, based on short handwriting tasks, YOLO26 improves global accuracy from 55% to 61% and patient-class F1-score from 52% to 57% with respect to YOLO11, but recall remains limited at 50%. In the long-writing setting, restricted to Task 14 and Task 25, both models reach 59% accuracy; however, their clinical behavior diverges sharply. YOLO11 attains 75% precision but only 33% patient recall, whereas YOLO26 reaches 72% recall and 65% F1-score. These findings show that task complexity is not a secondary variable: sustained sentence-copying tasks appear to expose disease-related graphomotor alterations more clearly than short traces. The study should be interpreted as a proof of concept, because the cohort is small, image-level reporting is used, and no external validation is available.

Handwriting Analysis,Alzheimer's Disease,Digital Biomarkers,Computer Vision,Patient-Level Split,YOLO11,YOLO26,DARWIN-I