Ayat Elmyerigiy, Emad Sami Jaha, Alaa Bafail · Diagnostics 2026 · 2026
DOI: 10.3390/diagnostics16223735
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Background/Objectives: Deep learning models can score the Clock Drawing Test (CDT) accurately, but their explanations are usually judged by looking at saliency maps. We asked whether complementary quantitative explainability criteria, evaluated against independent human annotations, give convergent evidence that a CDT model has learned interpretable drawing abnormalities. Methods: An image-only ResNet50 was trained to predict five-level CDT scores from the National Health and Aging Trends Study (7708 observations; participant-level split). In 200 separate, participant-unique drawings annotated for seven abnormalities, we evaluated Grad-CAM attribution alignment, matched-ink perturbation sensitivity, and linear probing of frozen representations. Sensitivity analyses covered region-of-interest definitions, data quality, and adjustment for CDT score. Results: On 1181 test observations, the quadratic weighted kappa was 0.740 and accuracy 0.678. Unadjusted analyses suggested abnormality-specific effects, including reduced hand-region attribution for missing or incorrect hands (q < 0.001) and significant perturbation sensitivity for three abnormalities. However, perturbation effects were as large in drawings without the abnormality. Within score levels, neither attribution nor perturbation sensitivity distinguished drawings with and without abnormalities. Within-score linear decodability remained for five abnormalities (AUROC 0.64–0.93), but not for uneven number spacing or same-length hands. Inter-rater agreement varied (κ 0.27–0.85). Conclusions: In this model and dataset, attribution and perturbation sensitivity mainly reflected overall drawing quality and the presence of intact hands and numerals. Abnormality-specific information was detectable only in the learned representation. Feature-level explainability evaluations should control for the prediction target, and saliency maps alone do not show that a model has learned interpretable abnormalities.
No comments yet — start the discussion below.